import json import os import re import time from dataclasses import dataclass, field from datetime import date from typing import Any, Dict, List, Optional, Set, Tuple, Union import gradio as gr import requests from bs4 import BeautifulSoup from duckduckgo_search import DDGS from huggingface_hub import InferenceClient # --- Model configuration --------------------------------------------------- # Our own DeepResearch model. When QUEST_BASE_URL is configured in Space # Secrets, the app will route requests to that dedicated HF Inference Endpoint # instead of the shared HF Inference API. QUEST_MODEL_ID = "osunlp/QUEST-35B" QUEST_BASE_URL = os.getenv("QUEST_BASE_URL", "").strip() # Endpoints built from the TGI image expose a single-model OpenAI route; the # model name passed to chat_completion is usually "tgi". vLLM endpoints usually # want the original repo id. QUEST_ENDPOINT_MODEL overrides this if needed. QUEST_ENDPOINT_MODEL = os.getenv("QUEST_ENDPOINT_MODEL", "tgi").strip() or "tgi" # This Space runs exclusively on QUEST-35B served via the private HF Inference # Endpoint pointed to by QUEST_BASE_URL. No public fallback list — the model # field in the UI is display-only. DEFAULT_MODEL = QUEST_MODEL_ID # Internal defaults. Search budget is no longer user-tunable. DEFAULT_MAX_SEARCH_RESULTS = 10 PAPER_URL = os.getenv("PAPER_URL", "https://arxiv.org/abs/2605.24218") CODE_URL = os.getenv("CODE_URL", "https://github.com/OSU-NLP-Group/QUEST") DATASET_URL = os.getenv("DATASET_URL", "https://huggingface.co/collections/osunlp/quest") MODEL_URL = os.getenv("MODEL_URL", "https://huggingface.co/osunlp/QUEST-35B-RL") # --- System prompt --------------------------------------------------------- # Full QUEST SYSTEM_PROMPT (mirrors inference/prompt.py in the research repo) # so that QUEST-35B sees the exact tool schema it was trained with. Other # models still follow this schema just fine in practice. QUEST_SYSTEM_PROMPT = """You are a deep research assistant. Your core function is to conduct thorough, multi-source investigations into any topic. You must handle both broad, open-domain inquiries and queries within specialized academic fields. For every request, synthesize information from credible, diverse sources to deliver a comprehensive, accurate, and objective response. When you have gathered sufficient information and are ready to provide the definitive response, you must enclose the entire final answer within tags. # Tools You may call one or more functions to assist with the user query. You are provided with function signatures within XML tags: {"type": "function", "function": {"name": "search", "description": "Perform Google web searches then returns a string of the top search results. Accepts multiple queries.", "parameters": {"type": "object", "properties": {"query": {"type": "array", "items": {"type": "string", "description": "The search query."}, "minItems": 1, "description": "The list of search queries."}}, "required": ["query"]}}} {"type": "function", "function": {"name": "visit", "description": "Visit webpage(s) and return the summary of the content.", "parameters": {"type": "object", "properties": {"url": {"type": "array", "items": {"type": "string"}, "description": "The URL(s) of the webpage(s) to visit. Can be a single URL or an array of URLs."}, "goal": {"type": "string", "description": "The specific information goal for visiting webpage(s)."}}, "required": ["url", "goal"]}}} {"type": "function", "function": {"name": "google_scholar", "description": "Leverage Google Scholar to retrieve relevant information from academic publications. Accepts multiple queries.", "parameters": {"type": "object", "properties": {"query": {"type": "array", "items": {"type": "string", "description": "The search query."}, "minItems": 1, "description": "The list of search queries for Google Scholar."}}, "required": ["query"]}}} # Using prev_state (Research State Summary) If you see a "RESEARCH STATE SUMMARY (prev_state)" section in the user message, it contains a compressed summary of previous research progress. Use it to: 1. **Avoid redundant work**: - Check `search_queries` to avoid repeating searches that have already been executed. - Check `visited_sources` to avoid visiting URLs that have already been visited. 2. **Use verified information**: - Check `information_state.trusted` for facts that have been verified from visited sources. You can use these directly in your answer without re-searching or re-visiting. - Check `information_state.untrusted` for claims that have been contradicted or proven unreliable. 3. **Follow up on uncertain information**: - Check `information_state.uncertain` for claims that need more evidence. The `need` field specifies the exact next action (e.g., "visit " or "search ") to resolve the uncertainty. IMPORTANT: Do NOT search for or visit information that is already in `prev_state`, unless it's insufficient to answer the user's question. Only in this case, you are encouraged to search for more information or even visit the same URL. Instead, use the information from `prev_state` directly, or follow the specific actions suggested in `information_state.uncertain.need` if more information is needed. The final answer must exclude any information that remains uncertain or pending. All statements included must be fully verified. For each function call, return a json object with function name and arguments within XML tags: {"name": , "arguments": } Current date: """ # --------------------------------------------------------------------------- # Vendored prompts from OSU-NLP-Group/QUEST (inference/prompt.py and # inference/tool_memory.py). Kept verbatim so the secondary-LLM behaviour # (visit extractor + condenser State Summarizer) matches the research code. # --------------------------------------------------------------------------- EXTRACTOR_PROMPT = """Please process the following webpage content and user goal to extract relevant information: ## **Webpage Content** {webpage_content} ## **User Goal** {goal} ## **Task Guidelines** 1. **Content Scanning for Rationale**: Locate the **specific sections/data** directly related to the user's goal within the webpage content 2. **Key Extraction for Evidence**: Identify and extract the **most relevant information** from the content, you never miss any important information, output the **full original context** of the content as far as possible, it can be more than three paragraphs. 3. **Summary Output for Summary**: Organize into a concise paragraph with logical flow, prioritizing clarity and judge the contribution of the information to the goal. **Final Output Format using JSON format has "rational", "evidence", "summary" feilds** """ MEMORY_SYSTEM_PROMPT = """You are a State Summarizer for a DeepResearch agent. Your ONLY job is to maintain a compact, parseable, context-aware state JSON for memory management. Your primary objective is to prevent redundant search and redundant visit actions by extracting useful, answer-ready information from tool responses and preserving it in a structured state. You will be given: 1) events: a chronological list of interaction events (user/assistant messages and tool calls/responses) 2) prev_state: the previous state JSON (may be empty or null) You MUST output ONLY a single JSON object that conforms EXACTLY to the schema below. No markdown, no extra text, no code fences, no explanations. ======================== OUTPUT JSON SCHEMA (STRICT) { "version": "dr_state", "search_queries": [ { "q": "string", "intent": "string" } ], "visited_sources": [ { "url": "string", "note": "string" } ], "information_state": { "trusted": [ { "id": "T1", "claim": "string", "sources": ["string"], "reason": "string" } ], "untrusted": [ { "id": "U1", "claim": "string", "sources": ["string"], "reason": "string" } ], "uncertain": [ { "id": "C1", "claim": "string", "sources": ["string"], "reason": "string", "need": "string" } ] } } ======================== TRIGGER NOTE (IMPORTANT) This summarizer is invoked automatically when CONTEXT_THRESHOLD is reached: - The system invokes summarization when context tokens reach a threshold. - Focus on extracting evidence, deduplicating tool usage, and making the state more actionable. Note: Agent-initiated condenser tool calls are ignored for memory updates. Only automatic CONTEXT_THRESHOLD triggers will update the memory state. ======================== CORE PRINCIPLE (CRITICAL) Visited pages alone are NOT useful memory. For every visit() tool_response, you MUST attempt to extract at least one useful, concrete fact into information_state unless the page is irrelevant. The goal is that the DeepResearch agent can rely on information_state.trusted to answer questions directly, and rely on information_state.uncertain.need to know the exact next step without re-searching. ======================== UPDATE RULES (IMPORTANT) 0) Anti-redundancy objective: - The state must clearly encode: a) what is already verified and final (trusted), b) what is false or contradicted (untrusted), c) what is missing AND the exact next action to resolve it (uncertain.need). - Prefer concrete actions such as: "visit " or "search ". 1) Merge with prev_state: - Start from prev_state if provided; update it using new events. - Never delete past entries except for: a) exact duplicates, or b) bucket migration (moving the same claim between uncertain/trusted/untrusted). 2) De-duplication: - search_queries: dedupe by exact "q" string. - visited_sources: dedupe by exact "url". - information_state: dedupe by exact "claim" string ACROSS ALL BUCKETS with priority: trusted > untrusted > uncertain. 3) Output ONLY the JSON object. No markdown, no extra text. Return ONLY the updated JSON object.""" def build_system_prompt() -> str: return QUEST_SYSTEM_PROMPT + date.today().isoformat() TOOL_RESPONSE_TEMPLATE = """ {payload} """ SEARCH_CACHE: Dict[str, Dict[str, Any]] = {} VISIT_CACHE: Dict[str, Dict[str, Any]] = {} # Quest paper palette. The Gradio shell is themed to match the OSU-NLP Quest # microsite: soft off-white page, paper-white cards, terracotta accent, mint # secondary, Manrope for UI type and Source Serif 4 for display headings. APP_THEME = gr.themes.Base( primary_hue=gr.themes.colors.orange, secondary_hue=gr.themes.colors.teal, neutral_hue=gr.themes.colors.slate, font=[ gr.themes.GoogleFont("Manrope"), "ui-sans-serif", "system-ui", "sans-serif", ], font_mono=[ gr.themes.GoogleFont("JetBrains Mono"), "ui-monospace", "monospace", ], ).set( body_background_fill="#F2F4F8", body_text_color="#0D1117", body_text_color_subdued="#64748B", color_accent="#BE5B2B", color_accent_soft="rgba(190,91,43,0.09)", background_fill_primary="#FFFFFF", background_fill_secondary="#EEF1F7", border_color_primary="rgba(10,15,40,0.08)", border_color_accent="#BE5B2B", block_background_fill="#FFFFFF", block_border_width="1px", block_border_color="rgba(10,15,40,0.08)", block_shadow="0 1px 2px rgba(10,15,40,0.05), 0 2px 10px rgba(10,15,40,0.06)", block_radius="16px", block_label_background_fill="transparent", block_label_border_width="0px", block_label_text_color="#64748B", block_label_text_weight="700", block_title_text_color="#0D1117", block_title_text_weight="700", block_title_border_width="0px", panel_background_fill="transparent", panel_border_width="0px", panel_border_color="transparent", input_background_fill="#FFFFFF", input_background_fill_focus="#FFFFFF", input_border_color="rgba(10,15,40,0.12)", input_border_color_focus="#BE5B2B", input_border_width="1px", input_radius="12px", input_shadow="none", input_shadow_focus="0 0 0 3px rgba(190,91,43,0.15)", code_background_fill="#EEF1F7", slider_color="#BE5B2B", button_primary_background_fill="#0D1117", button_primary_background_fill_hover="#1F2A37", button_primary_text_color="#FFFFFF", button_primary_border_color="transparent", button_primary_shadow="0 1px 2px rgba(10,15,40,0.08), 0 6px 18px rgba(10,15,40,0.12)", button_secondary_background_fill="#FFFFFF", button_secondary_background_fill_hover="rgba(190,91,43,0.09)", button_secondary_text_color="#BE5B2B", button_secondary_border_color="rgba(10,15,40,0.16)", button_cancel_background_fill="#FFFFFF", button_cancel_background_fill_hover="#FEE2E2", button_cancel_text_color="#DC2626", button_cancel_border_color="#FCA5A5", table_border_color="rgba(10,15,40,0.08)", table_even_background_fill="#FAFBFD", table_odd_background_fill="#FFFFFF", ) CUSTOM_CSS = """ /* === Quest paper palette applied to the Gradio shell ==================== */ /* Brings the OSU-NLP Quest microsite aesthetic into the live Space: soft off-white background, paper-white cards with subtle 1px borders and low-opacity shadows, terracotta accent, Source Serif 4 for display headings, Manrope for everything else. */ :root { --q-bg: #F2F4F8; --q-paper: #FFFFFF; --q-surface-alt: #EEF1F7; --q-line: rgba(10, 15, 40, 0.08); --q-line-strong: rgba(10, 15, 40, 0.16); --q-text: #0D1117; --q-muted: #64748B; --q-accent: #BE5B2B; --q-accent-soft: rgba(190, 91, 43, 0.09); --q-accent-line: rgba(190, 91, 43, 0.55); --q-mint: #0B9E8A; --q-mint-deep: #0A8070; --q-cover-bg: #0D1117; --q-shadow: 0 1px 3px rgba(10,15,40,0.04), 0 8px 32px rgba(10,15,40,0.08); --q-shadow-card: 0 1px 2px rgba(10,15,40,0.05), 0 2px 10px rgba(10,15,40,0.06); --q-radius-xl: 20px; --q-radius-lg: 16px; --q-radius-md: 12px; } html, body, gradio-app, [class*="gradio-container"] { background: var(--q-bg) !important; } /* Full-height shell ------------------------------------------------------- */ html, body { width: 100% !important; min-height: 100vh !important; margin: 0 !important; font-size: 17px !important; } gradio-app { display: block !important; width: 100% !important; min-height: 100vh !important; margin-left: auto !important; margin-right: auto !important; } gradio-app > .gradio-container, gradio-app > div { display: block !important; width: 100% !important; margin-left: auto !important; margin-right: auto !important; } [class*="gradio-container"] { max-width: 1700px !important; width: 100% !important; min-width: 320px !important; margin-left: auto !important; margin-right: auto !important; padding: 28px 36px 72px !important; color: var(--q-text); box-sizing: border-box !important; font-family: "Manrope", ui-sans-serif, system-ui, sans-serif; font-size: 1rem !important; } [class*="gradio-container"] *::selection { background: rgba(190,91,43,0.18); } /* Prevent inner wrappers from collapsing when streaming content first arrives. */ [class*="gradio-container"] .layout-gap { width: 100% !important; } [class*="gradio-container"] .layout-gap > .gr-column, [class*="gradio-container"] .layout-gap > div { min-width: 0 !important; } [class*="gradio-container"] .gradio-markdown, [class*="gradio-container"] [data-testid="markdown"] { min-height: 220px !important; } [class*="gradio-container"] .codemirror-wrapper, [class*="gradio-container"] .cm-editor { min-height: 220px !important; } /* Long code / markdown cannot push the layout sideways. */ [class*="gradio-container"] .gradio-code, [class*="gradio-container"] .gradio-markdown, [class*="gradio-container"] .prose, [class*="gradio-container"] .markdown, [class*="gradio-container"] [data-testid="markdown"], [class*="gradio-container"] .tabs, [class*="gradio-container"] .tabitem, [class*="gradio-container"] .tab-content { max-width: 100% !important; width: 100% !important; min-width: 0 !important; word-wrap: break-word !important; overflow-wrap: anywhere !important; } [class*="gradio-container"] .codemirror-wrapper { max-width: 100% !important; border-radius: 14px !important; overflow: hidden !important; } [class*="gradio-container"] .cm-editor { max-width: 100% !important; overflow: hidden !important; } [class*="gradio-container"] .cm-scroller { max-width: 100% !important; overflow-x: auto !important; } [class*="gradio-container"] .cm-content, [class*="gradio-container"] .cm-line { max-width: 100% !important; white-space: pre-wrap !important; word-break: break-word !important; } [class*="gradio-container"] .prose pre, [class*="gradio-container"] .markdown pre { max-width: 100% !important; overflow-x: auto !important; white-space: pre-wrap !important; } /* === Quest-style header ================================================= */ .quest-header { display: flex; align-items: center; gap: 18px; padding: 18px 22px; margin: 8px 0 24px; border: 1px solid var(--q-line); border-radius: var(--q-radius-lg); background: var(--q-paper); box-shadow: var(--q-shadow-card); } .quest-header-mark { display: grid; place-items: center; width: 48px; height: 48px; flex-shrink: 0; border-radius: 12px; background: var(--q-text); color: #FFFFFF; font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif; font-weight: 700; font-size: 1.55rem; } .quest-header-text { display: flex; flex-direction: column; gap: 4px; min-width: 0; } .quest-header-title { margin: 0; font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif; font-weight: 600; font-size: clamp(1.1rem, 1.8vw, 1.5rem); line-height: 1.2; letter-spacing: -0.01em; color: var(--q-text); } .quest-header-byline { color: var(--q-muted); font-size: 0.9rem; font-weight: 500; text-decoration: underline; text-decoration-color: rgba(100,116,139,0.45); text-underline-offset: 3px; text-decoration-thickness: 1px; width: fit-content; transition: color 140ms ease, text-decoration-color 140ms ease; } .quest-header-byline:hover { color: var(--q-accent); text-decoration-color: var(--q-accent); } /* === Cards (section-card) =============================================== */ .section-card { background: var(--q-paper) !important; border: 1px solid var(--q-line) !important; border-radius: var(--q-radius-xl) !important; box-shadow: var(--q-shadow-card) !important; padding: 22px !important; } .no-frame { background: transparent !important; border: none !important; box-shadow: none !important; padding: 0 !important; } /* Section kicker + hero heading follow the paper treatment. */ .section-heading { font-size: 0.7rem; font-weight: 800; letter-spacing: 0.14em; text-transform: uppercase; color: var(--q-accent); margin: 0 0 14px 0; } .section-heading-row { display: flex; align-items: baseline; justify-content: space-between; gap: 16px; flex-wrap: wrap; } .hero-heading { font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif !important; font-weight: 600 !important; /* Kept identical to .quest-header-title so the two display headings match in size and line spacing. */ font-size: clamp(1.1rem, 1.8vw, 1.5rem) !important; line-height: 1.2 !important; letter-spacing: -0.01em !important; text-transform: none !important; color: var(--q-text) !important; } /* Match the .brand mark from the Quest microsite (github-page branch). */ .quest-name { font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif !important; font-style: italic !important; font-weight: 700 !important; color: inherit !important; letter-spacing: -0.005em; margin: 4px 0 14px 0 !important; } .hero-subtitle { color: var(--q-muted); font-size: 0.95rem; line-height: 1.6; margin: -6px 0 16px 0; } /* Layout gap: mirror the paper's column rhythm. */ .layout-gap { gap: 24px !important; align-items: flex-start; } .right-stack > * { margin-bottom: 14px; } .action-row { gap: 10px !important; margin-top: 14px; } .action-row button { min-width: 0; flex: 1; } /* === Icon grid (Paper / Code / Dataset / Model) ========================= */ .icon-grid { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 10px; width: 100%; } .icon-link { display: flex; align-items: center; justify-content: center; gap: 8px; padding: 11px 14px; border-radius: 999px; text-decoration: none !important; color: var(--q-text) !important; background: var(--q-paper); font-weight: 600; font-size: 0.88rem; white-space: nowrap; border: 1px solid var(--q-line-strong); transition: background 140ms ease, border-color 140ms ease, color 140ms ease, transform 140ms ease; } .icon-link:hover { background: var(--q-accent-soft); border-color: var(--q-accent-line); color: var(--q-accent) !important; transform: translateY(-1px); } /* Resource cards (paper / code / data / model) — icon + label, eye-catching */ .resource-grid { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 10px; width: 100%; } .resource-card { display: flex; align-items: center; gap: 10px; padding: 12px 14px; border-radius: 14px; text-decoration: none !important; color: var(--q-text) !important; background: var(--q-paper); border: 1px solid var(--q-line-strong); transition: background 140ms ease, border-color 140ms ease, color 140ms ease, transform 140ms ease; } .resource-card:hover { background: var(--q-accent-soft); border-color: var(--q-accent-line); color: var(--q-accent) !important; transform: translateY(-1px); } .resource-card-icon { display: inline-flex; align-items: center; justify-content: center; width: 30px; height: 30px; flex-shrink: 0; border-radius: 8px; background: var(--q-surface-alt); color: var(--q-text); } .resource-card-icon svg { width: 18px; height: 18px; fill: currentColor; } .resource-card-icon.resource-card-emoji { background: transparent; font-size: 22px; line-height: 1; } .resource-card-text { display: flex; flex-direction: column; line-height: 1.15; min-width: 0; } .resource-card-text strong { font-weight: 700; font-size: 0.92rem; } .resource-card-text small { font-size: 0.72rem; color: var(--q-muted); margin-top: 2px; } /* === Buttons ============================================================ */ [class*="gradio-container"] button.primary, [class*="gradio-container"] .gr-button-primary { background: var(--q-text) !important; color: #ffffff !important; border: 1px solid var(--q-text) !important; box-shadow: 0 1px 2px rgba(10,15,40,0.08), 0 6px 18px rgba(10,15,40,0.12) !important; font-weight: 700 !important; letter-spacing: 0.01em !important; } [class*="gradio-container"] button.primary:hover, [class*="gradio-container"] .gr-button-primary:hover { background: #1F2A37 !important; border-color: #1F2A37 !important; } [class*="gradio-container"] button.secondary, [class*="gradio-container"] .gr-button-secondary { background: var(--q-paper) !important; color: var(--q-text) !important; border: 1px solid var(--q-line-strong) !important; box-shadow: none !important; font-weight: 600 !important; } [class*="gradio-container"] button.secondary:hover, [class*="gradio-container"] .gr-button-secondary:hover { background: var(--q-accent-soft) !important; border-color: var(--q-accent-line) !important; color: var(--q-accent) !important; } [class*="gradio-container"] button.stop, [class*="gradio-container"] .gr-button-stop { background: var(--q-paper) !important; color: #DC2626 !important; border: 1px solid #FCA5A5 !important; box-shadow: none !important; font-weight: 600 !important; } [class*="gradio-container"] button.stop:hover, [class*="gradio-container"] .gr-button-stop:hover { background: #FEE2E2 !important; color: #B91C1C !important; } /* Flatten every grey block Gradio drops inside our cards. */ [class*="gradio-container"] .gr-group, [class*="gradio-container"] fieldset, [class*="gradio-container"] .gr-box, [class*="gradio-container"] .gr-panel, [class*="gradio-container"] .form, [class*="gradio-container"] .gr-form, [class*="gradio-container"] .container { background: transparent !important; } .section-card { --block-shadow: none; --block-shadow-dark: none; --block-background-fill: transparent; --block-border-color: transparent; --block-border-width: 0px; --panel-background-fill: transparent; --panel-border-width: 0px; --background-fill-secondary: transparent; --border-color-primary: transparent; overflow: visible !important; } .section-card > div, .section-card > div > div, .section-card > div > div > div { background: transparent !important; border: none !important; box-shadow: none !important; overflow: visible !important; } .section-card .block, .section-card .form, .section-card .gr-form, .section-card .gr-block, .section-card .gr-panel, .section-card .gr-group, .section-card .gradio-dropdown, .section-card .gradio-slider, .section-card .gradio-textbox, .section-card .gradio-markdown, .section-card .gradio-code { background: transparent !important; border: none !important; box-shadow: none !important; overflow: visible !important; } .section-card .form, .section-card .gr-form { display: flex !important; flex-direction: column !important; gap: 14px !important; } [class*="gradio-container"] .section-card .row, [class*="gradio-container"] .section-card [class*="row"] { display: flex !important; flex-direction: row !important; flex-wrap: wrap !important; gap: 10px !important; } .action-row { display: flex !important; flex-direction: row !important; gap: 10px !important; margin-top: 14px; } .action-row > * { flex: 1 1 0; min-width: 0; } .section-card > * + * { margin-top: 14px; } /* === Inputs ============================================================= */ [class*="gradio-container"] textarea, [class*="gradio-container"] input:not([type="checkbox"]):not([type="radio"]):not([type="range"]) { background: var(--q-paper) !important; border: 1px solid var(--q-line-strong) !important; box-shadow: none !important; border-radius: var(--q-radius-md) !important; color: var(--q-text) !important; font-family: "Manrope", ui-sans-serif, system-ui, sans-serif !important; } /* Make the Model Textbox match the Memory Strategy Dropdown's height (46px outer = 44px content + 2*1px border). */ .section-card [data-testid="textbox"] textarea, .section-card [data-testid="textbox"] input { min-height: 44px !important; padding: 11px 14px !important; line-height: 1.4 !important; box-sizing: border-box !important; } [class*="gradio-container"] textarea::placeholder, [class*="gradio-container"] input::placeholder { color: #94A3B8 !important; } [class*="gradio-container"] textarea:focus, [class*="gradio-container"] input:focus { border-color: var(--q-accent) !important; box-shadow: 0 0 0 3px rgba(190,91,43,0.15) !important; outline: none !important; } /* === Dropdown =========================================================== */ [class*="gradio-container"] [data-testid="dropdown"], [class*="gradio-container"] .gradio-dropdown { background: var(--q-paper) !important; border: 1px solid var(--q-line-strong) !important; border-radius: var(--q-radius-md) !important; box-shadow: none !important; padding: 0 !important; min-height: 46px !important; width: 100% !important; box-sizing: border-box !important; } [class*="gradio-container"] [data-testid="dropdown"] > .wrap, [class*="gradio-container"] [data-testid="dropdown"] .secondary-wrap, [class*="gradio-container"] [data-testid="dropdown"] .wrap-inner, [class*="gradio-container"] [data-testid="dropdown"] .input-container, [class*="gradio-container"] [data-testid="dropdown"] .single-select, [class*="gradio-container"] .gradio-dropdown .wrap, [class*="gradio-container"] .gradio-dropdown .wrap-inner, [class*="gradio-container"] .gradio-dropdown .secondary-wrap, [class*="gradio-container"] .gradio-dropdown .input-container, [class*="gradio-container"] .gradio-dropdown .single-select, [class*="gradio-container"] [class*="dropdown"] .wrap { background: transparent !important; border: 0 !important; outline: 0 !important; box-shadow: none !important; border-radius: 0 !important; width: 100% !important; min-height: 44px !important; padding: 0 14px !important; display: flex !important; align-items: center !important; box-sizing: border-box !important; } [class*="gradio-container"] [data-testid="dropdown"] input, [class*="gradio-container"] .gradio-dropdown input, [class*="gradio-container"] [data-testid="dropdown"] select, [class*="gradio-container"] .gradio-dropdown select { background: transparent !important; border: 0 !important; outline: 0 !important; box-shadow: none !important; padding: 0 !important; height: 44px !important; line-height: 44px !important; font-size: 0.95rem !important; width: 100% !important; border-radius: 0 !important; } /* Force-remove any nested pill/rounded background that makes the dropdown look like it has two concentric frames. */ [class*="gradio-container"] [data-testid="dropdown"] .container, [class*="gradio-container"] [data-testid="dropdown"] .wrap > .wrap, [class*="gradio-container"] .gradio-dropdown .container, [class*="gradio-container"] .gradio-dropdown .wrap > .wrap { border: 0 !important; outline: 0 !important; box-shadow: none !important; background: transparent !important; border-radius: 0 !important; padding: 0 !important; } /* The little caret/arrow icon container — vertically center it */ [class*="gradio-container"] [data-testid="dropdown"] .icon-wrap, [class*="gradio-container"] .gradio-dropdown .icon-wrap { top: 50% !important; transform: translateY(-50%) !important; right: 14px !important; } [class*="gradio-container"] .options ul, [class*="gradio-container"] .options { background: var(--q-paper) !important; border: 1px solid var(--q-line) !important; border-radius: var(--q-radius-md) !important; box-shadow: 0 10px 30px rgba(10,15,40,0.12) !important; } [class*="gradio-container"] .options li[aria-selected="true"], [class*="gradio-container"] .options li:hover { background: var(--q-accent-soft) !important; color: var(--q-accent) !important; } /* Info hint text under inputs */ [class*="gradio-container"] .info, [class*="gradio-container"] [data-testid*="info"], [class*="gradio-container"] .gr-info { color: var(--q-muted) !important; background: transparent !important; font-size: 12px !important; } /* === Sliders ============================================================ */ /* Flatten the Slider's outer wrapper — Gradio paints a rectangular block around the label + track + value-input by default; remove it. */ .section-card .gradio-slider, .section-card .gradio-slider > div, .section-card .gradio-slider .form, .section-card .gradio-slider .gr-form, .section-card .gradio-slider .wrap, .section-card .gradio-slider .container, .section-card .gradio-slider .head { background: transparent !important; border: 0 !important; box-shadow: none !important; padding: 0 !important; } /* === Per-component flatteners (id-based; max specificity vs Gradio defaults) === */ /* The Memory Strategy dropdown and the two sliders ship with an outer block wrapper that paints a small rectangle. Flatten the wrapper AND any nested div Gradio inserts (form/container/wrap/etc), keeping label + interactive element visible. */ #quest-memory-strategy, #quest-memory-strategy > div, #quest-memory-strategy .form, #quest-memory-strategy .gr-form, #quest-memory-strategy .container, #quest-memory-strategy .wrap-inner, #quest-memory-strategy .head, #quest-max-turns, #quest-max-turns > div, #quest-max-turns .form, #quest-max-turns .gr-form, #quest-max-turns .container, #quest-max-turns .wrap-inner, #quest-max-turns .head, #quest-temperature, #quest-temperature > div, #quest-temperature .form, #quest-temperature .gr-form, #quest-temperature .container, #quest-temperature .wrap-inner, #quest-temperature .head, #quest-model, #quest-model > div, #quest-model .form, #quest-model .gr-form, #quest-model .container, #quest-model .wrap-inner, #quest-model .head { background: transparent !important; border: 0 !important; outline: 0 !important; box-shadow: none !important; padding: 0 !important; border-radius: 0 !important; } /* Memory Strategy radio: stack vertically, terracotta-tinted check state. */ #quest-memory-strategy .wrap, #quest-memory-strategy fieldset, #quest-memory-strategy [data-testid="radio"] { display: flex !important; flex-direction: column !important; gap: 6px !important; background: transparent !important; border: 0 !important; padding: 0 !important; } #quest-memory-strategy label { background: transparent !important; border: 1px solid var(--q-line) !important; border-radius: 8px !important; padding: 8px 12px !important; cursor: pointer !important; font-weight: 500 !important; font-size: 0.95rem !important; color: var(--q-text) !important; text-transform: none !important; letter-spacing: 0 !important; display: flex !important; align-items: center !important; gap: 10px !important; transition: border-color 120ms ease, background 120ms ease; } #quest-memory-strategy label:hover { border-color: var(--q-line-strong) !important; } #quest-memory-strategy input[type="radio"] { accent-color: var(--q-accent) !important; width: 16px !important; height: 16px !important; } /* Slider head input (the "[6 ↺]" / "[1 ↺]" pill next to the slider track): the global input rule paints a 1px border on it, which looks like a stray rectangle. Flatten it AND hide the reset button (it's redundant — the slider's range already shows the default value). */ #quest-max-turns input[type="number"], #quest-temperature input[type="number"] { border: 0 !important; background: transparent !important; box-shadow: none !important; border-radius: 0 !important; padding: 0 !important; min-height: 0 !important; height: auto !important; text-align: center !important; width: 3.5em !important; font-weight: 600 !important; color: var(--q-text) !important; } #quest-max-turns button, #quest-temperature button { display: none !important; } [class*="gradio-container"] input[type="range"] { -webkit-appearance: none; appearance: none; width: 100%; height: 6px; background: var(--q-surface-alt); border-radius: 999px; outline: none; box-shadow: none !important; border: none !important; } [class*="gradio-container"] input[type="range"]::-webkit-slider-runnable-track { height: 6px; background: linear-gradient(90deg,var(--q-accent) var(--val,50%), var(--q-surface-alt) var(--val,50%)); border-radius: 999px; } [class*="gradio-container"] input[type="range"]::-webkit-slider-thumb { -webkit-appearance: none; appearance: none; width: 18px; height: 18px; border-radius: 50%; background: #ffffff; border: 2px solid var(--q-accent); box-shadow: 0 2px 6px rgba(190,91,43,0.25); margin-top: -6px; cursor: pointer; } [class*="gradio-container"] input[type="range"]::-moz-range-track { height: 6px; background: var(--q-surface-alt); border-radius: 999px; } [class*="gradio-container"] input[type="range"]::-moz-range-progress { height: 6px; background: var(--q-accent); border-radius: 999px; } [class*="gradio-container"] input[type="range"]::-moz-range-thumb { width: 16px; height: 16px; border-radius: 50%; background: #ffffff; border: 2px solid var(--q-accent); box-shadow: 0 2px 6px rgba(190,91,43,0.25); } /* === Tabs =============================================================== */ [class*="gradio-container"] .tabs, [class*="gradio-container"] .tab-container, [class*="gradio-container"] .tab-wrapper { background: transparent !important; } [class*="gradio-container"] .tab-container::after { background: var(--q-line) !important; } [class*="gradio-container"] .tab-wrapper button { color: var(--q-muted) !important; font-weight: 700 !important; letter-spacing: 0.04em !important; text-transform: uppercase !important; font-size: 0.78rem !important; } [class*="gradio-container"] .tab-wrapper button.selected { color: var(--q-accent) !important; } [class*="gradio-container"] .tab-wrapper button.selected::after { background: var(--q-accent) !important; } /* Hide the orange streaming-progress bar that Gradio paints at the top of the Markdown/Code panel while a run is in flight. */ [class*="gradio-container"] .progress, [class*="gradio-container"] .progress-level, [class*="gradio-container"] .progress-level-inner, [class*="gradio-container"] .progress-bar, [class*="gradio-container"] .progress-text, [class*="gradio-container"] [class*="progress-level"], [class*="gradio-container"] .generating, [class*="gradio-container"] div[class*="progress-bar"] { display: none !important; background: transparent !important; border: 0 !important; height: 0 !important; } /* Kill any stray orange/thick separator that Gradio paints above the tab panel content (border-top or ::before on the tab content wrapper). */ [class*="gradio-container"] .tabitem, [class*="gradio-container"] .tab-content, [class*="gradio-container"] .gradio-tabitem, [class*="gradio-container"] .tabs > div.tabitem { border-top: 0 !important; box-shadow: none !important; background: transparent !important; } [class*="gradio-container"] .tabitem::before, [class*="gradio-container"] .tab-content::before, [class*="gradio-container"] .gradio-tabitem::before { content: none !important; } [class*="gradio-container"] .tab-nav, [class*="gradio-container"] .tab-wrapper { border-bottom: 1px solid var(--q-line) !important; border-top: 0 !important; } [class*="gradio-container"] .tab-nav::before, [class*="gradio-container"] .tab-wrapper::before { content: none !important; } /* Block labels above components */ [class*="gradio-container"] .gr-block label, [class*="gradio-container"] .gradio-slider label, [class*="gradio-container"] .gradio-dropdown label, [class*="gradio-container"] .gradio-textbox label { color: var(--q-muted) !important; font-weight: 700 !important; font-size: 0.74rem !important; letter-spacing: 0.08em !important; text-transform: uppercase !important; } /* === Markdown / prose =================================================== */ [class*="gradio-container"] .gr-markdown, [class*="gradio-container"] .prose, [class*="gradio-container"] .markdown { color: var(--q-text) !important; font-family: "Manrope", ui-sans-serif, system-ui, sans-serif !important; line-height: 1.75; } [class*="gradio-container"] .gr-markdown a, [class*="gradio-container"] .prose a { color: var(--q-accent) !important; text-decoration: underline; text-decoration-color: rgba(190,91,43,0.35); } [class*="gradio-container"] .gr-markdown a:hover, [class*="gradio-container"] .prose a:hover { text-decoration-color: var(--q-accent); } [class*="gradio-container"] .gr-markdown h1, [class*="gradio-container"] .gr-markdown h2, [class*="gradio-container"] .gr-markdown h3, [class*="gradio-container"] .prose h1, [class*="gradio-container"] .prose h2, [class*="gradio-container"] .prose h3 { font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif !important; font-weight: 600 !important; letter-spacing: -0.01em !important; color: var(--q-text) !important; } [class*="gradio-container"] .gr-markdown code, [class*="gradio-container"] .prose code { background: var(--q-surface-alt); border: 1px solid var(--q-line); padding: 1px 6px; border-radius: 6px; font-size: 0.9em; } /* === Code block (Record tab) ============================================ */ [class*="gradio-container"] .codemirror-wrapper, [class*="gradio-container"] .cm-editor, [class*="gradio-container"] .cm-scroller, [class*="gradio-container"] .cm-gutters, [class*="gradio-container"] .cm-content { background: var(--q-surface-alt) !important; color: var(--q-text) !important; border: none !important; font-family: "JetBrains Mono", ui-monospace, monospace !important; } [class*="gradio-container"] .cm-gutters { border-right: 1px solid var(--q-line) !important; color: var(--q-muted) !important; } /* === Rounded corners on everything ====================================== */ [class*="gradio-container"] .block, [class*="gradio-container"] .form, [class*="gradio-container"] .gr-box, [class*="gradio-container"] .gr-panel, [class*="gradio-container"] .gr-group, [class*="gradio-container"] [data-testid="textbox"], [class*="gradio-container"] [data-testid="dropdown"], [class*="gradio-container"] .tabitem, [class*="gradio-container"] .tab-content, [class*="gradio-container"] .gradio-markdown, [class*="gradio-container"] .gradio-code { border-radius: var(--q-radius-md) !important; } [class*="gradio-container"] button { border-radius: 999px !important; } /* === Example buttons ==================================================== */ .example-note { color: var(--q-muted); font-size: 13px; margin: 0 0 12px 0; line-height: 1.5; } .connection-tip { /* Reset the uppercase / accent inherited from .section-heading so it reads like a normal hint when placed inline next to a heading. */ color: var(--q-muted); font-size: 12px; font-weight: 500; letter-spacing: 0; text-transform: none; line-height: 1.45; } .memory-help { color: var(--q-muted); font-size: 12.5px; line-height: 1.55; margin: 6px 0 0 0; padding: 10px 12px; background: var(--q-surface-alt); border: 1px solid var(--q-line); border-radius: 8px; } .memory-help b { color: var(--q-text); font-weight: 600; } .example-buttons { display: grid; gap: 10px; margin-top: 4px; } [class*="gradio-container"] .example-btn { text-align: left !important; justify-content: flex-start !important; white-space: normal !important; line-height: 1.5 !important; padding: 14px 16px !important; font-size: 14px !important; color: var(--q-text) !important; background: var(--q-paper) !important; border: 1px solid var(--q-line) !important; border-radius: var(--q-radius-md) !important; box-shadow: none !important; font-weight: 500 !important; letter-spacing: normal !important; text-transform: none !important; } [class*="gradio-container"] .example-btn:hover { background: var(--q-accent-soft) !important; border-color: var(--q-accent-line) !important; color: var(--q-accent) !important; } [class*="gradio-container"] .example-btn > * { color: inherit !important; white-space: normal !important; display: inline !important; } /* Footer tagline block */ .quest-footer { margin-top: 28px; padding: 18px 24px; border: 1px solid var(--q-line); border-radius: var(--q-radius-xl); background: var(--q-paper); box-shadow: var(--q-shadow-card); display: flex; align-items: center; justify-content: space-between; gap: 20px; color: var(--q-muted); font-size: 0.86rem; line-height: 1.65; } .quest-footer a { color: var(--q-muted); text-decoration: none; } .quest-footer a:hover { color: var(--q-text); } .quest-footer-links { display: flex; gap: 16px; flex-wrap: wrap; } /* Tiny mark that replaces the HF watermark block. */ footer { display: none !important; } /* === Responsive ========================================================= */ @media (max-width: 1100px) { .quest-cover-inner { grid-template-columns: 1fr; } .quest-cover-panel.wide { grid-column: auto; min-height: 180px; } } @media (max-width: 760px) { [class*="gradio-container"] { padding: 16px !important; } .quest-footer { flex-direction: column; align-items: flex-start; } } /* Mobile / tablet: stack the two-column (Ask the agent | Open release) layout vertically. The columns are designed for ~420px + ~340px side by side; below ~900px they get squished — and because an earlier rule forces `min-width: 0` on them, they squeeze instead of wrapping, collapsing text to one character per line. Forcing the row to flex-direction:column makes each column take the full width and stack cleanly. NOTE: Gradio 5.x renders the row as `.row.layout-gap` and the columns as `.column` (no `gradio-container` ancestor, no `.gr-column`), so the selectors must be plain. */ @media (max-width: 900px) { .layout-gap, .row.layout-gap { flex-direction: column !important; flex-wrap: wrap !important; } .layout-gap > .column, .layout-gap > div { width: 100% !important; max-width: 100% !important; flex: 1 1 auto !important; } /* Action buttons (Run Research / Stop / Clear): three equal flex:1 columns don't leave room for the two-word "Run Research" label on a phone, so it squished into a circle with overflowing text. Stack them full-width. */ .action-row { flex-wrap: wrap !important; gap: 8px !important; } .action-row > * { flex: 1 1 100% !important; min-width: 100% !important; } .action-row button { white-space: nowrap !important; border-radius: 12px !important; } /* Open-release cards: one per row on mobile. The 2-up grid squeezes each card to ~87px, char-wrapping the Paper / Code / Data / Model labels. */ .resource-grid { grid-template-columns: 1fr !important; } } /* Phone-specific: shrink the two big display headings so they don't dominate the small screen. Both get the SAME size + line-height so they stay visually consistent with each other. */ @media (max-width: 600px) { .quest-header-title, .hero-heading { font-size: 0.9rem !important; line-height: 1.2 !important; } } """ @dataclass class AgentState: searched_queries: List[str] = field(default_factory=list) visited_urls: List[str] = field(default_factory=list) searched_query_set: Set[str] = field(default_factory=set) visited_url_set: Set[str] = field(default_factory=set) trusted_notes: List[str] = field(default_factory=list) trace: List[Dict[str, Any]] = field(default_factory=list) # Accept a variety of placeholder-only answers: a bare ellipsis (ASCII `...` # or unicode `…`), a single interpunct, and any whitespace-only content. These # show up when the model echoes a literal `...` template # from the prompt instead of producing a real answer. _PLACEHOLDER_ANSWER_RE = re.compile(r"^[\s.\u2026\u00b7]*$") # Pipe-table separator line, e.g. `| --- | :---: |`. The outer pipes are # optional in some GFM dialects, so we accept both. _TABLE_SEPARATOR_RE = re.compile( r"^\s*\|?\s*:?-{2,}:?(?:\s*\|\s*:?-{2,}:?)+\s*\|?\s*$" ) def strip_think_blocks(text: str) -> str: """Remove any ... reasoning blocks. QUEST-35B (Qwen3 family) emits `` reasoning before the final answer. When the endpoint is deployed without a reasoning parser, the raw tags leak into chat completion `content`; stripping them here keeps the extracted answer clean for Markdown rendering. """ return re.sub( r".*?", "", text, flags=re.DOTALL | re.IGNORECASE ) def decode_escaped_whitespace(text: str) -> str: """Decode literal `\\n`/`\\t`/`\\r` sequences back to real whitespace. Some OpenAI-compatible servers (and some vLLM builds when a tokenizer's chat template escapes control characters) return `choices[0].message.content` with newlines stored as the two-character backslash+n sequence rather than as a real newline. That breaks Markdown rendering because a pipe table on a single line is not a table — it is just a sentence with `|` in it, which is exactly the symptom we saw with: \\n| Color | Hex |\\n|---|---|\\n| Red | #FF0000 |... We only decode when the escapes dominate (at least 3 of them, and at least as many as the real newlines in the text). That keeps us from corrupting legitimate backslash-n pairs that happen to appear in a code sample the model produced. """ if not text: return text escaped_newlines = text.count("\\n") if escaped_newlines == 0 and "\\t" not in text and "\\r" not in text: return text real_newlines = text.count("\n") if escaped_newlines < max(3, real_newlines + 1): return text # Preserve real backslashes so that `\\\\n` (an actual `\n` the model # wrote) doesn't get collapsed to a newline. sentinel = "\x00__BS__\x00" out = text.replace("\\\\", sentinel) out = out.replace("\\n", "\n").replace("\\r", "\r").replace("\\t", "\t") out = out.replace(sentinel, "\\") return out def _is_placeholder_answer(text: str) -> bool: return bool(_PLACEHOLDER_ANSWER_RE.match(text or "")) def ensure_markdown_table_blank_lines(text: str) -> str: """Insert a blank line before any pipe-table header row. GitHub-Flavored Markdown requires a pipe table to be preceded by a paragraph break; otherwise the header row is folded into the previous paragraph and the whole table renders as raw text. Models sometimes glue the table directly under a sentence (e.g. "Here's the comparison: | Col ..."), so we fix that up defensively. """ lines = text.split("\n") out: List[str] = [] for idx, line in enumerate(lines): is_header = ( "|" in line and idx + 1 < len(lines) and _TABLE_SEPARATOR_RE.match(lines[idx + 1]) is not None ) if is_header and out and out[-1].strip() != "": out.append("") out.append(line) return "\n".join(out) def extract_answer(text: str) -> Optional[str]: """Return the content of the first `...` block. Tries two strategies, in order, and discards placeholder-only content (bare ellipses) that the model sometimes echoes from the prompt: 1. Well-formed `...` block. 2. Truncated `...` with no closing tag (tokens ran out); in that case we take everything after the opening tag. """ # Decode escaped whitespace on the whole output first so the # regex can actually match the opening and closing tags across lines. decoded = decode_escaped_whitespace(text or "") cleaned = strip_think_blocks(decoded) full_match = re.search( r"\s*(.*?)\s*", cleaned, flags=re.DOTALL | re.IGNORECASE, ) if full_match is not None: candidate = decode_escaped_whitespace(full_match.group(1).strip()) if candidate and not _is_placeholder_answer(candidate): return candidate # Closed block was a placeholder / empty: fail fast. Do NOT fall # through to the open-ended strategy, or it would re-match the same # tag and incorrectly capture `...` as the answer. return None open_match = re.search( r"\s*(.*)$", cleaned, flags=re.DOTALL | re.IGNORECASE ) if open_match is not None: candidate = decode_escaped_whitespace(open_match.group(1).strip()) if candidate and not _is_placeholder_answer(candidate): return candidate return None def parse_tool_call(text: str) -> Tuple[Optional[str], Optional[Dict[str, Any]], Optional[str]]: cleaned = strip_think_blocks(text or "") match = re.search(r"\s*(.*?)\s*", cleaned, flags=re.DOTALL | re.IGNORECASE) if not match: return None, None, None payload = match.group(1).strip() try: data = json.loads(payload) except json.JSONDecodeError: return None, None, "Invalid JSON in block." name = data.get("name") arguments = data.get("arguments", {}) if not isinstance(name, str) or not isinstance(arguments, dict): return None, None, "Invalid tool format. Expect name(str) and arguments(dict)." return name, arguments, None _SEARCH_UNAVAILABLE_HINT = ( "The web-search backend is currently rate-limited or unreachable. " "If this question can be answered confidently from your own training " "knowledge (e.g. common product specs, historical facts, definitions), " "please produce your best answer now inside ..., and " "mention any value that might be out of date. Only ask the user to " "retry later if the question truly requires a fresh web lookup." ) # Google Serper API key. Either SERPER_API_KEY or SERPER_KEY_ID is accepted # so that the Space matches the env-var name used by the research repo. SERPER_API_KEY = ( os.getenv("SERPER_API_KEY") or os.getenv("SERPER_KEY_ID") or "" ).strip() SERPER_ENDPOINT = os.getenv("SERPER_ENDPOINT", "https://google.serper.dev/search") def _serper_search(query: str, max_results: int) -> Dict[str, Any]: """Hit the Google Serper API. Returns the same shape as `_ddg_search`. Serper responds in well under a second and is not subject to the 202 Ratelimit we get from html.duckduckgo.com, so preferring it when the key is set cuts latency dramatically and eliminates most search failures on shared Space IPs. """ try: resp = requests.post( SERPER_ENDPOINT, headers={ "X-API-KEY": SERPER_API_KEY, "Content-Type": "application/json", }, json={"q": query, "num": max_results}, timeout=15, ) resp.raise_for_status() data = resp.json() except Exception as exc: return { "ok": False, "query": query, "error": f"Serper error: {type(exc).__name__}: {exc}", "results": [], "backend": "serper", } rows: List[Dict[str, str]] = [] for item in (data.get("organic") or [])[:max_results]: rows.append( { "title": item.get("title", ""), "href": item.get("link", ""), "body": item.get("snippet", ""), } ) # Fold in the answer box, knowledge graph and "people also ask" when # present; for many queries `organic` is empty but these rich blocks # still carry the exact fact the model is looking for. answer_box = data.get("answerBox") or {} if answer_box: rows.insert( 0, { "title": answer_box.get("title", "Answer box"), "href": answer_box.get("link", ""), "body": answer_box.get("snippet") or answer_box.get("answer") or "", }, ) kg = data.get("knowledgeGraph") or {} if kg: kg_body = kg.get("description", "") or "" attrs = kg.get("attributes") or {} if attrs: kg_body = (kg_body + " " + "; ".join( f"{k}: {v}" for k, v in attrs.items() )).strip() if kg.get("title") or kg_body: rows.append( { "title": kg.get("title", "Knowledge graph"), "href": kg.get("descriptionLink") or kg.get("website", ""), "body": kg_body, } ) for paa in (data.get("peopleAlsoAsk") or [])[:3]: if paa.get("question"): rows.append( { "title": paa.get("question", ""), "href": paa.get("link", ""), "body": paa.get("snippet", ""), } ) if not rows: # Serper worked but the query genuinely matched nothing. Return a # SUCCESSFUL-but-empty result (ok=True) carrying an actionable hint, # rather than ok=False — that way the agent is told to reformulate # instead of seeing a scary "all backends failed" message, and we # skip the rate-limited DuckDuckGo fallback (which would also find # nothing). Mirrors inference/tool_search.py's "use a less specific # query" behaviour. return { "ok": True, "query": query, "results": [], "cached": False, "backend": "serper", "note": ( "No results for this query. Reformulate and search again: " "remove quotation marks (they force exact-phrase matching), " "drop rare acronyms, and use fewer / more common keywords." ), } return { "ok": True, "query": query, "results": rows, "cached": False, "backend": "serper", } def _ddg_search(query: str, max_results: int) -> Dict[str, Any]: """Fallback path: scrape DuckDuckGo. Rate-limits on shared IPs.""" last_exc: Optional[BaseException] = None for attempt in range(2): try: rows: List[Dict[str, str]] = [] with DDGS() as ddgs: for item in ddgs.text(query, max_results=max_results): rows.append( { "title": item.get("title", ""), "href": item.get("href", ""), "body": item.get("body", ""), } ) return { "ok": True, "query": query, "results": rows, "cached": False, "backend": "duckduckgo", } except Exception as exc: last_exc = exc if attempt == 0: time.sleep(1.5) continue err = f"{type(last_exc).__name__}: {last_exc}" if last_exc else "unknown error" return { "ok": False, "query": query, "error": f"DuckDuckGo unavailable ({err}).", "results": [], "backend": "duckduckgo", } def _run_search_single(query: str, max_results: int) -> Dict[str, Any]: """Run one search query, preferring Serper when the key is set. Returns a structured dict on both success and failure; never raises. Order of preference: 1. Google Serper (fast, no scraping, requires `SERPER_API_KEY` / `SERPER_KEY_ID`). 2. DuckDuckGo HTML backend (free, but rate-limits on shared Space IPs). 3. Graceful `ok: False` payload with a hint that tells the agent to answer from its own knowledge if it reasonably can. """ if not query.strip(): return {"ok": False, "error": "Search query cannot be empty."} cache_key = f"{query.strip().lower()}::{max_results}" if cache_key in SEARCH_CACHE: return {**SEARCH_CACHE[cache_key], "cached": True} tried: List[Dict[str, Any]] = [] if SERPER_API_KEY: serper_result = _serper_search(query, max_results) if serper_result.get("ok"): SEARCH_CACHE[cache_key] = serper_result return serper_result tried.append(serper_result) ddg_result = _ddg_search(query, max_results) if ddg_result.get("ok"): SEARCH_CACHE[cache_key] = ddg_result return ddg_result tried.append(ddg_result) # Both backends failed (or no Serper key and DDG rate-limited). errors = "; ".join( f"{r.get('backend', 'unknown')}: {r.get('error', 'no results')}" for r in tried ) return { "ok": False, "query": query, "error": f"All search backends failed ({errors}).", "results": [], "hint": _SEARCH_UNAVAILABLE_HINT, } def run_search(query: Union[str, List[str]], max_results: int = 5) -> Dict[str, Any]: """Runs one or more queries through DuckDuckGo. QUEST's schema passes `query` as an array of strings, while the simpler starter schema used a single string. We accept both shapes. """ if isinstance(query, list): sub_results: List[Dict[str, Any]] = [] for q in query: if not isinstance(q, str) or not q.strip(): continue sub_results.append(_run_search_single(q, max_results)) return {"ok": True, "queries": query, "results": sub_results} return _run_search_single(str(query or "").strip(), max_results) def _clean_html_to_text(html: str, max_chars: int) -> str: soup = BeautifulSoup(html, "html.parser") for tag in soup(["script", "style", "noscript"]): tag.decompose() text = soup.get_text(separator=" ", strip=True) text = re.sub(r"\s+", " ", text) return text[:max_chars] # --------------------------------------------------------------------------- # Secondary-LLM helpers (visit extractor, condenser State Summarizer, scholar). # Mirror inference/tool_visit.py + inference/tool_memory.py + inference/tool_scholar.py. # Each helper is best-effort: if the relevant env vars are missing it returns # None / falls through to the legacy behaviour so the Space still works. # --------------------------------------------------------------------------- JINA_API_KEYS = os.getenv("JINA_API_KEYS", "").strip() WEBCONTENT_MAXLENGTH = int(os.getenv("WEBCONTENT_MAXLENGTH", "60000")) SUMMARY_MODEL_NAME = os.getenv("SUMMARY_MODEL_NAME", "").strip() SUMMARY_API_KEY = (os.getenv("API_KEY") or os.getenv("SUMMARY_OPENAI_API_KEY") or "").strip() SUMMARY_API_BASE = (os.getenv("API_BASE") or os.getenv("SUMMARY_OPENAI_BASE_URL") or "").strip() or None MEMORY_MODEL_NAME = os.getenv("MEMORY_MODEL_NAME", "").strip() MEMORY_API_KEY = (os.getenv("MEMORY_OPENAI_API_KEY") or SUMMARY_API_KEY).strip() MEMORY_API_BASE = (os.getenv("MEMORY_OPENAI_BASE_URL") or SUMMARY_API_BASE) or None MEMORY_TOKEN_THRESHOLD = int( os.getenv("MEMORY_THRESHOLD") or os.getenv("MEMORY_CONTEXT_THRESHOLD") or os.getenv("MEMORY_TOKEN_THRESHOLD") or "80000" ) # Azure OpenAI support — mirrors inference/tool_visit.py logic. When # AZURE_OPENAI_ENDPOINT is set, we use AzureOpenAI() instead of OpenAI() # and AZURE_OPENAI_DEPLOYMENT overrides the per-purpose model name. AZURE_OPENAI_ENDPOINT = os.getenv("AZURE_OPENAI_ENDPOINT", "").strip() AZURE_OPENAI_API_VERSION = os.getenv("AZURE_OPENAI_API_VERSION", "").strip() or "2024-06-01" AZURE_OPENAI_DEPLOYMENT = os.getenv("AZURE_OPENAI_DEPLOYMENT", "").strip() def _get_chat_client_and_model( api_key: str, base_url: Optional[str], fallback_model_name: str ) -> Tuple[Optional[Any], str]: """Construct an OpenAI-compatible chat client. Auto-switches to AzureOpenAI when AZURE_OPENAI_ENDPOINT is configured; in that case the effective model name becomes AZURE_OPENAI_DEPLOYMENT (Azure uses deployment names, not raw model ids). Returns (client, model_name).""" if AZURE_OPENAI_ENDPOINT: try: from openai import AzureOpenAI except Exception: return None, fallback_model_name if not api_key: return None, fallback_model_name client = AzureOpenAI( api_key=api_key, azure_endpoint=AZURE_OPENAI_ENDPOINT, api_version=AZURE_OPENAI_API_VERSION, ) return client, (AZURE_OPENAI_DEPLOYMENT or fallback_model_name) try: from openai import OpenAI except Exception: return None, fallback_model_name if not api_key: return None, fallback_model_name client = ( OpenAI(api_key=api_key, base_url=base_url) if base_url else OpenAI(api_key=api_key) ) return client, fallback_model_name # Backwards-compat shim: older callers asked for just a client without # Azure awareness. Keep returning the non-Azure client so we don't break # anything if a future patch imports it. def _get_openai_client(api_key: str, base_url: Optional[str]): client, _ = _get_chat_client_and_model(api_key, base_url, fallback_model_name="") return client def _approx_token_count(text: str) -> int: """Cheap token estimate (~4 chars/token). Tiktoken is heavy; this is fine for threshold gating where being off by 20% is harmless.""" try: import tiktoken return len(tiktoken.get_encoding("cl100k_base").encode(text)) except Exception: return max(1, len(text) // 4) def _messages_token_count(messages: List[Dict[str, str]]) -> int: return sum(_approx_token_count(str(m.get("content", ""))) for m in messages) def _jina_readpage(url: str) -> Optional[str]: """Fetch a page via Jina Reader (r.jina.ai). Returns markdown text on success, None on failure (caller falls back to BeautifulSoup).""" if not JINA_API_KEYS: return None headers = {"Authorization": f"Bearer {JINA_API_KEYS}"} for attempt in range(3): try: r = requests.get(f"https://r.jina.ai/{url}", headers=headers, timeout=50) if r.status_code == 200 and r.text: return r.text[:WEBCONTENT_MAXLENGTH] except Exception: if attempt == 2: return None return None # Last error from the most recent _llm_extract call (for diagnostics in trace). _LAST_EXTRACT_ERR: Optional[str] = None def _llm_extract(webpage_content: str, goal: str) -> Optional[str]: """Run the SUMMARY model as the visit extractor. Mirrors inference/prompt.py:build_visit_extractor_messages + tool_visit's call. Picks AzureOpenAI when AZURE_OPENAI_ENDPOINT is set.""" global _LAST_EXTRACT_ERR _LAST_EXTRACT_ERR = None if not SUMMARY_API_KEY: _LAST_EXTRACT_ERR = "API_KEY env var not set" return None client, model_name = _get_chat_client_and_model( SUMMARY_API_KEY, SUMMARY_API_BASE, SUMMARY_MODEL_NAME ) if client is None: _LAST_EXTRACT_ERR = "openai client could not be constructed (package missing?)" return None if not model_name: _LAST_EXTRACT_ERR = ( "no model name (set SUMMARY_MODEL_NAME or, on Azure, " "AZURE_OPENAI_DEPLOYMENT)" ) return None try: resp = client.chat.completions.create( model=model_name, messages=[ { "role": "user", "content": EXTRACTOR_PROMPT.format( webpage_content=webpage_content, goal=goal or "general overview" ), } ], timeout=120, ) return (resp.choices[0].message.content or "").strip() or None except Exception as exc: # Show underlying cause too — OpenAI SDK's APIConnectionError wraps # the real httpx/ssl/dns error which is what we usually need to see. cause = repr(exc.__cause__) if getattr(exc, "__cause__", None) else "no cause" _LAST_EXTRACT_ERR = ( f"{type(exc).__name__}: {str(exc)[:200]} | cause={cause[:200]} | " f"endpoint={AZURE_OPENAI_ENDPOINT or SUMMARY_API_BASE or 'default'} | " f"model={model_name} | azure={'yes' if AZURE_OPENAI_ENDPOINT else 'no'}" ) return None def _llm_condense(events_text: str, prev_state: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]: """Run the MEMORY model as the State Summarizer. Returns a parsed JSON state dict, or None if condensation failed. Picks AzureOpenAI when AZURE_OPENAI_ENDPOINT is set.""" client, model_name = _get_chat_client_and_model( MEMORY_API_KEY, MEMORY_API_BASE, MEMORY_MODEL_NAME ) if client is None or not model_name: return None user_payload = json.dumps( { "events": events_text[-30000:], # cap input "prev_state": prev_state or None, }, ensure_ascii=False, ) try: resp = client.chat.completions.create( model=model_name, messages=[ {"role": "system", "content": MEMORY_SYSTEM_PROMPT}, {"role": "user", "content": user_payload}, ], timeout=180, ) raw = (resp.choices[0].message.content or "").strip() # the prompt says no code fences, but be defensive anyway if raw.startswith("```"): raw = re.sub(r"^```(?:json)?\s*|\s*```$", "", raw, flags=re.DOTALL) return json.loads(raw) except Exception: return None def _run_scholar_single(query: str) -> Dict[str, Any]: """Google Scholar via Serper. Mirrors inference/tool_scholar.py.""" q = (query or "").strip() if not q: return {"ok": False, "error": "Scholar query cannot be empty."} if not SERPER_API_KEY: return { "ok": False, "query": q, "error": "SERPER_API_KEY missing — scholar tool unavailable.", } headers = {"X-API-KEY": SERPER_API_KEY, "Content-Type": "application/json"} payload = json.dumps({"q": q}) last_err: Optional[str] = None for _ in range(3): try: r = requests.post( "https://google.serper.dev/scholar", data=payload, headers=headers, timeout=20, ) if r.status_code == 200: data = r.json() rows = [] for page in data.get("organic", []) or []: rows.append( { "title": page.get("title", ""), "link": page.get("link", ""), "year": page.get("year"), "publicationInfo": page.get("publicationInfo"), "snippet": page.get("snippet", ""), "citedBy": page.get("citedBy"), } ) return {"ok": True, "query": q, "results": rows, "backend": "serper-scholar"} last_err = f"HTTP {r.status_code}: {r.text[:200]}" except Exception as exc: last_err = f"{type(exc).__name__}: {exc}" return {"ok": False, "query": q, "error": f"Serper scholar failed ({last_err})."} def _run_visit_single(url: str, max_chars: int, goal: str = "") -> Dict[str, Any]: if not url.strip(): return {"ok": False, "error": "URL cannot be empty."} cache_key = f"{url.strip()}::{max_chars}::{goal[:60]}" if cache_key in VISIT_CACHE: return {**VISIT_CACHE[cache_key], "cached": True, "goal": goal} # Preferred path: Jina Reader for clean markdown → LLM extractor distils # the page content against the requested goal. Matches the research repo's # inference/tool_visit.py behaviour. Either step failing falls through to # the legacy requests + BeautifulSoup path. jina_md = _jina_readpage(url) if jina_md: extract = _llm_extract(jina_md, goal) if SUMMARY_MODEL_NAME else None result = { "ok": True, "url": url, "goal": goal, "content": (extract or jina_md)[:max_chars], "extractor": "llm" if extract else "jina-raw", } if not extract and _LAST_EXTRACT_ERR: result["extractor_error"] = _LAST_EXTRACT_ERR VISIT_CACHE[cache_key] = result return result try: resp = requests.get( url, timeout=20, headers={"User-Agent": "Mozilla/5.0 (compatible; DeepResearchSpace/1.0)"}, ) resp.raise_for_status() content_type = resp.headers.get("content-type", "") if "text/html" in content_type or " Dict[str, Any]: """Fetches one or more URLs. Accepts string or list (QUEST schema).""" if isinstance(url, list): sub_results: List[Dict[str, Any]] = [] for u in url: if not isinstance(u, str) or not u.strip(): continue sub_results.append(_run_visit_single(u, max_chars, goal)) return {"ok": True, "goal": goal, "results": sub_results} return _run_visit_single(str(url or "").strip(), max_chars, goal) def _build_client_for_model(model: str) -> Tuple[InferenceClient, str, List[str]]: """Returns (client, primary_model_id, fallback_model_ids). When the user picks the Quest model and QUEST_BASE_URL is configured, the InferenceClient is pointed at the dedicated endpoint; otherwise we hit the shared HF Inference API and let the starter fall back across free models. """ token = os.getenv("HF_TOKEN") quest_timeout = int(os.getenv("QUEST_REQUEST_TIMEOUT", "600")) if model == QUEST_MODEL_ID and QUEST_BASE_URL: # Prefer a dedicated key for the self-hosted endpoint so the real HF # token never travels into vLLM / tunnel logs. endpoint_token = os.getenv("QUEST_API_KEY") or token client = InferenceClient( base_url=QUEST_BASE_URL, token=endpoint_token, timeout=quest_timeout, ) return client, QUEST_ENDPOINT_MODEL, [] client = InferenceClient(token=token, timeout=quest_timeout) return client, model, [] def call_model( client: InferenceClient, messages: List[Dict[str, str]], preferred_model: str, candidate_models: List[str], temperature: float, max_new_tokens: int, ) -> Tuple[str, str]: model_order: List[str] = [] for m in [preferred_model] + candidate_models: if m and m not in model_order: model_order.append(m) last_error = None for model_name in model_order: try: completion = client.chat_completion( model=model_name, messages=messages, temperature=temperature, max_tokens=max_new_tokens, ) return completion.choices[0].message.content or "", model_name except Exception as exc: last_error = exc continue raise RuntimeError(f"All model candidates failed. Last error: {last_error}") def _render_progress( lines: List[str], used_model: str, question: str, ) -> str: """Render the in-progress status view that replaces the Markdown panel while the agent is still running, so the user is not staring at a blank box for the 20-60 seconds a full QUEST-35B research run can take.""" header = ( f"### ⏳ Researching…\n\n" f"**Model:** `{used_model}` \n" f"**Question:** {question.strip()[:200]}" ) if not lines: body = "_Starting agent…_" else: body = "\n".join(f"- {line}" for line in lines) return f"{header}\n\n{body}" def _trace_to_json(state: "AgentState", used_model: str) -> str: return json.dumps( { "used_model": used_model, "searched_queries": state.searched_queries, "visited_urls": state.visited_urls, "trusted_notes": state.trusted_notes[-10:], "trace": state.trace, }, ensure_ascii=False, indent=2, ) MEMORY_STRATEGIES = ("condenser", "vanilla", "discard_all", "hide_tool_result") def _normalize_memory_strategy(strategy: str) -> str: s = (strategy or "condenser").strip().lower().replace("-", "_") if s == "hide_tool_results": s = "hide_tool_result" return s if s in MEMORY_STRATEGIES else "condenser" def _apply_memory_strategy(messages: List[Dict[str, str]], strategy: str, turn: int) -> None: """Lightweight port of the strategies defined in the Quest inference code (`inference/react_agent.py`). Upstream is token-threshold-driven; this Space approximates each strategy on a turn-count basis for demo purposes. - vanilla: no-op (matches MEMORY_ENABLED=false upstream). - condenser: no-op here; the main loop injects a compact research-state summary every few turns (a poor-man's stand-in for the upstream State Summarizer LLM that emits a structured trusted/untrusted/ uncertain JSON when the token threshold is hit). - discard_all: every 8 turns, reset history to [system, user question] (upstream resets when token_count crosses the threshold). - hide_tool_result: keep only the most recent tool-response user message; older ones get their content replaced with a stub (mirrors upstream behavior). """ if strategy == "discard_all": if turn > 1 and turn % 8 == 0 and len(messages) > 2: system_msg = messages[0] question_msg = messages[1] messages.clear() messages.append(system_msg) messages.append(question_msg) messages.append( { "role": "user", "content": "[memory discarded at turn " f"{turn} — continue the research from the original question]", } ) elif strategy == "hide_tool_result": keep_tail = 1 tool_indices = [ i for i, m in enumerate(messages) if m.get("role") == "user" and str(m.get("content", "")).startswith("") ] if len(tool_indices) > keep_tail: for i in tool_indices[:-keep_tail]: if messages[i]["content"] != "[hidden]": messages[i] = { "role": "user", "content": "[hidden]", } def build_research_agent( question: str, model: str, max_turns: int, temperature: float, memory_strategy: str = "condenser", ): """Run the ReAct research loop as a generator. Each `yield` emits a `(markdown_for_answer_panel, json_for_record_panel)` tuple. Intermediate yields show progress so that Gradio streams the status lines into the UI as work happens. The last yield contains the final answer and the final trace. """ client, primary_model, fallback_models = _build_client_for_model(model) # Display label: the real HF repo id is nicer than the TGI shim name. display_primary = model if (model == QUEST_MODEL_ID) else primary_model state = AgentState() used_model = display_primary status_lines: List[str] = [] def _emit(): """Yield the current progress snapshot to Gradio.""" return ( _render_progress(status_lines, used_model, question), _trace_to_json(state, used_model), ) messages: List[Dict[str, str]] = [ {"role": "system", "content": build_system_prompt()}, {"role": "user", "content": question}, ] final_answer: Optional[str] = None # `prev_state` holds the JSON returned by the State Summarizer LLM. It is # refreshed each time the context tokens cross MEMORY_TOKEN_THRESHOLD and # then injected into the model's next user message as a RESEARCH STATE # SUMMARY block. Matches inference/react_agent.py + inference/tool_memory.py # behaviour. prev_state: Optional[Dict[str, Any]] = None condenser_runs = 0 status_lines.append("🚀 Starting research agent") yield _emit() strategy = _normalize_memory_strategy(memory_strategy) os.environ["MEMORY_STRATEGY"] = strategy for turn in range(1, max_turns + 1): _apply_memory_strategy(messages, strategy, turn) # Real LLM-based condenser: when tokens cross the threshold, call the # MEMORY model to produce the structured state JSON, then rebuild the # context as [system, original_question, RESEARCH_STATE_SUMMARY]. if ( strategy == "condenser" and (MEMORY_MODEL_NAME or AZURE_OPENAI_DEPLOYMENT) and MEMORY_API_KEY and turn > 1 and _messages_token_count(messages) > MEMORY_TOKEN_THRESHOLD ): status_lines.append( f"🗜️ turn {turn}: condensing context (tokens > {MEMORY_TOKEN_THRESHOLD})" ) yield _emit() events_text = "\n\n".join( f"[{m.get('role')}] {str(m.get('content',''))[:2000]}" for m in messages[2:] # skip system + original question ) new_state = _llm_condense(events_text, prev_state) if new_state: prev_state = new_state condenser_runs += 1 state.trace.append( {"turn": turn, "condenser_run": condenser_runs, "prev_state": prev_state} ) # Reset history to system + question + state summary summary_block = ( "RESEARCH STATE SUMMARY (prev_state)\n" + json.dumps(prev_state, ensure_ascii=False, indent=2) + "\n\nUse this summary to avoid redundant work and " "follow `information_state.uncertain.need` for next steps." ) messages[:] = [messages[0], messages[1], {"role": "user", "content": summary_block}] status_lines[-1] = ( f"🗜️ turn {turn}: condensed → " f"{len(prev_state.get('information_state', {}).get('trusted', []))} trusted, " f"{len(prev_state.get('information_state', {}).get('uncertain', []))} uncertain" ) yield _emit() elif ( strategy == "condenser" and not ((MEMORY_MODEL_NAME or AZURE_OPENAI_DEPLOYMENT) and MEMORY_API_KEY) and state.trusted_notes and turn > 1 and turn % 3 == 0 ): # Fallback heuristic when the MEMORY model is not configured. summary_lines = "\n".join(f"- {n}" for n in state.trusted_notes[-6:]) messages.append( { "role": "user", "content": f"RESEARCH STATE SUMMARY\n{summary_lines}\nUse this summary to avoid repeating work.", } ) status_lines.append(f"🧠 turn {turn}: thinking…") yield _emit() t0 = time.time() raw_output, endpoint_model = call_model( client=client, messages=messages, preferred_model=primary_model, candidate_models=fallback_models, temperature=temperature, max_new_tokens=int(os.getenv("QUEST_MAX_NEW_TOKENS", "4096")), ) dt = time.time() - t0 model_output = raw_output # Preserve the human-friendly model id for the trace even if the # endpoint ignores the "model" param and returns the TGI shim name. used_model = display_primary if endpoint_model == primary_model == QUEST_ENDPOINT_MODEL else endpoint_model messages.append({"role": "assistant", "content": model_output}) state.trace.append({"turn": turn, "assistant": model_output, "elapsed_s": round(dt, 2)}) status_lines[-1] = f"🧠 turn {turn}: model reply in {dt:.1f}s" yield _emit() extracted_answer = extract_answer(model_output) if extracted_answer: final_answer = extracted_answer status_lines.append("✍️ writing final answer") yield _emit() break tool_name, tool_args, tool_err = parse_tool_call(model_output) if tool_err: tool_response = {"ok": False, "error": tool_err} status_lines.append(f"⚠️ turn {turn}: malformed tool call — {tool_err}") yield _emit() elif not tool_name: # No explicit tool call and no final answer: force finalization. # IMPORTANT: do not write the literal characters `...` # here. Some models (notably the Qwen3 family that QUEST-35B is # built on) will echo the template verbatim, which means the # extracted answer ends up being the three-dot placeholder `...` # and the user sees an empty-looking result. messages.append( { "role": "user", "content": ( "You did not call a tool and did not produce a final " "answer. Please now write your best final answer, " "wrapped between an opening tag and a " "closing tag. Put the real answer text " "between those tags; do not write a literal ellipsis " "or other placeholder. If the question asks for " "tabular data, use GitHub-Flavored Markdown pipe " "tables (`| col1 | col2 |` + `|---|---|`) and put a " "blank line before the first row so the table renders." ), } ) status_lines.append(f"🙃 turn {turn}: model stalled; asking for an answer") yield _emit() continue else: if tool_name == "search": raw_query = tool_args.get("query", "") queries: List[str] if isinstance(raw_query, list): queries = [str(q).strip() for q in raw_query if str(q).strip()] else: queries = [str(raw_query).strip()] if str(raw_query).strip() else [] max_results = int(tool_args.get("max_results", DEFAULT_MAX_SEARCH_RESULTS)) max_results = max(1, min(max_results, DEFAULT_MAX_SEARCH_RESULTS)) queries_preview = ", ".join(f"`{q}`" for q in queries) or "_(empty)_" status_lines.append(f"🔍 turn {turn}: searching {queries_preview}") yield _emit() per_query: List[Dict[str, Any]] = [] backend_labels: List[str] = [] hits_total = 0 for q in queries: if q in state.searched_query_set: per_query.append({ "ok": True, "query": q, "cached": True, "note": "Already searched; reusing cached result.", "results": [], }) backend_labels.append("cache") continue state.searched_queries.append(q) state.searched_query_set.add(q) single = _run_search_single(q, max_results) per_query.append(single) backend_labels.append(single.get("backend", "unknown")) if single.get("ok"): hits_total += len(single.get("results", [])) first_titles = [r.get("title", "") for r in single.get("results", [])[:2]] if first_titles: state.trusted_notes.append( f"Searched '{q}' and found leads: {', '.join(t for t in first_titles if t)}" ) else: status_lines.append( f"⚠️ search failed on `{q}` via {single.get('backend', 'unknown')}: " f"{single.get('error', 'no results')}" ) tool_response = ( per_query[0] if len(per_query) == 1 else {"ok": True, "queries": queries, "results": per_query} ) unique_backends = sorted(set(backend_labels)) backend_str = "/".join(unique_backends) if unique_backends else "?" status_lines.append( f"✅ turn {turn}: got {hits_total} hit(s) via {backend_str}" ) yield _emit() elif tool_name == "visit": raw_url = tool_args.get("url", "") urls: List[str] if isinstance(raw_url, list): urls = [str(u).strip() for u in raw_url if str(u).strip()] else: urls = [str(raw_url).strip()] if str(raw_url).strip() else [] goal = str(tool_args.get("goal", "")).strip() max_chars = int(tool_args.get("max_chars", 6000)) max_chars = max(500, min(max_chars, 20000)) urls_preview = ", ".join(f"`{u[:60]}`" for u in urls) or "_(empty)_" status_lines.append(f"🌐 turn {turn}: visiting {urls_preview}") yield _emit() per_url: List[Dict[str, Any]] = [] visit_ok = 0 for u in urls: if u in state.visited_url_set: per_url.append({ "ok": True, "url": u, "cached": True, "note": "Already visited; reusing cached result.", }) visit_ok += 1 continue state.visited_urls.append(u) state.visited_url_set.add(u) single = _run_visit_single(u, max_chars, goal) per_url.append(single) if single.get("ok"): visit_ok += 1 snippet = str(single.get("content", ""))[:180] if snippet: state.trusted_notes.append( f"Visited {u} and extracted key context: {snippet}" ) tool_response = ( per_url[0] if len(per_url) == 1 else {"ok": True, "goal": goal, "results": per_url} ) status_lines.append( f"✅ turn {turn}: read {visit_ok}/{len(urls)} page(s)" ) yield _emit() elif tool_name in ("google_scholar", "scholar"): raw_query = tool_args.get("query", "") queries: List[str] if isinstance(raw_query, list): queries = [str(q).strip() for q in raw_query if str(q).strip()] else: queries = [str(raw_query).strip()] if str(raw_query).strip() else [] queries_preview = ", ".join(f"`{q}`" for q in queries) or "_(empty)_" status_lines.append(f"🎓 turn {turn}: scholar {queries_preview}") yield _emit() per_q = [_run_scholar_single(q) for q in queries] tool_response = ( per_q[0] if len(per_q) == 1 else {"ok": True, "results": per_q} ) ok_count = sum(1 for r in per_q if r.get("ok")) status_lines.append( f"📚 turn {turn}: scholar {ok_count}/{len(per_q)} ok" ) yield _emit() else: tool_response = {"ok": False, "error": f"Unknown tool: {tool_name}"} status_lines.append(f"⚠️ turn {turn}: unknown tool `{tool_name}`") yield _emit() state.trace.append({"turn": turn, "tool": tool_name, "tool_response": tool_response}) messages.append( { "role": "user", "content": TOOL_RESPONSE_TEMPLATE.format( payload=json.dumps(tool_response, ensure_ascii=False) ), } ) if final_answer is None: final_answer = ( "I could not finish a complete research answer within the configured turns. " "Try increasing max turns or switching to a stronger model." ) else: final_answer = ensure_markdown_table_blank_lines(final_answer) final_answer = f"**Model used:** `{used_model}`\n\n{final_answer}" trace_text = _trace_to_json(state, used_model) yield (final_answer, trace_text) def run_ui( question: str, max_turns: int, memory_strategy: str, temperature: float, ): if not question.strip(): yield "Please input a question.", "{}" return if not os.getenv("HF_TOKEN"): warning = ( "HF_TOKEN is not configured in Space Secrets. " "Go to Settings -> Secrets -> add `HF_TOKEN`, then retry." ) yield warning, json.dumps({"error": warning}, ensure_ascii=False, indent=2) return if not QUEST_BASE_URL: warning = ( f"`{QUEST_MODEL_ID}` needs a private HF Inference Endpoint. " "Create one at https://ui.endpoints.huggingface.co/, then set " "`QUEST_BASE_URL` in Space Secrets to the endpoint's `/v1/` URL." ) yield warning, json.dumps({"error": warning}, ensure_ascii=False, indent=2) return try: for partial_answer, partial_trace in build_research_agent( question=question, model=QUEST_MODEL_ID, max_turns=max_turns, temperature=temperature, memory_strategy=memory_strategy, ): yield partial_answer, partial_trace except Exception as exc: yield f"Error: {exc}", json.dumps({"error": str(exc)}, ensure_ascii=False, indent=2) EXAMPLES = [ { "category": "Multi-hop facts", "icon": "🎯", "text": "Who was the first person to walk on the Moon, and which U.S. President set that goal in his famous 1962 “Moon speech”?", }, { "category": "Time-varying + multi-hop", "icon": "📈", "text": "Who is the current CEO of the company that acquired GitHub in 2018, and what was that company's market capitalization at the close of the most recent quarter?", }, { "category": "Multi-constraint", "icon": "🧩", "text": "Find a 2-day itinerary in Tokyo under $250 focused on contemporary art museums and vegetarian restaurants, including transit between sites.", }, { "category": "Research Report", "icon": "📚", "text": "Compare the LLM-safety research approaches of Anthropic, OpenAI, and Google DeepMind over the past 18 months, focusing on alignment techniques and red-teaming methodologies.", }, ] def _example_label(ex: Dict[str, str]) -> str: return f"{ex['icon']} {ex['category']} — {ex['text']}" with gr.Blocks( title="QUEST · Deep Research by OSU NLP", theme=APP_THEME, css=CUSTOM_CSS, fill_width=True, ) as demo: # --- Quest-style header (Q mark + title + byline) --- gr.HTML( """

QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks

""" ) # --- Main two-column layout --- with gr.Row(elem_classes="layout-gap"): with gr.Column(scale=6, min_width=420): with gr.Group(elem_classes="section-card"): gr.HTML( '
Ask the agent
' '
QUEST: What can I research for you?
' ) question = gr.Textbox( show_label=False, placeholder="Ask anything you want to research in depth...", lines=6, ) with gr.Row(elem_classes="action-row"): run_btn = gr.Button("Run Research", variant="primary", size="lg") stop_btn = gr.Button("Stop", variant="stop", size="lg") clear_btn = gr.Button("Clear", variant="secondary", size="lg") with gr.Group(elem_classes="section-card"): gr.HTML( '
Try examples
' '
QUEST can handle multiple types of queries as shown below.
' ) with gr.Column(elem_classes="example-buttons"): example_buttons = [ gr.Button(_example_label(ex), variant="secondary", elem_classes="example-btn") for ex in EXAMPLES ] with gr.Group(elem_classes="section-card"): gr.HTML( '
' 'Output' '' 'If you see a connection error, please wait a moment and retry.' '' '
' ) with gr.Tabs(): with gr.TabItem("Result"): answer = gr.Markdown(label="Final Answer") with gr.TabItem("Record"): trace = gr.Code(label="Execution Trace (JSON)", language="json") with gr.Column(scale=4, min_width=340, elem_classes="right-stack"): with gr.Group(elem_classes="section-card"): gr.HTML( f"""
Open release
""" ) with gr.Group(elem_classes="section-card"): gr.HTML('
Settings
') gr.Textbox( label="Model", value=QUEST_MODEL_ID, interactive=False, elem_id="quest-model", ) memory_strategy = gr.Radio( label="Memory Strategy", choices=[ ("Condenser (default)", "condenser"), ("Vanilla", "vanilla"), ("Discard-all", "discard_all"), ("Hide-tool-result", "hide_tool_result"), ], value="condenser", elem_id="quest-memory-strategy", ) gr.HTML( '
' 'Condenser (default) — when context grows large, a State Summarizer LLM compresses earlier turns into a structured JSON of trusted/untrusted/uncertain claims, visited sources, and prior search queries; the agent continues with that compact state.
' 'Vanilla — memory management disabled; the full conversation history is kept.
' 'Discard-all — when context grows large, the entire message history is reset, restarting the agent from the original question with no accumulated context.
' 'Hide-tool-result — when context grows large, older tool responses are pruned; only the most recent tool result is kept.' '
' ) max_turns = gr.Slider( label="Max Turns", minimum=2, maximum=50, value=15, step=1, elem_id="quest-max-turns", ) temperature = gr.Slider( label="Temperature", minimum=0.0, maximum=1.5, value=1.0, step=0.1, elem_id="quest-temperature", ) gr.HTML( """ """ ) # IMPORTANT: keep `run_event` pointing at the .click() (the long-running # generator), not at the chained .then(). stop_btn.cancels=[run_event] # must target the generator for Stop to actually interrupt it; if we # capture the .then() result instead, Stop cancels the instant clear # lambda and the agent keeps running. run_event = run_btn.click( fn=run_ui, inputs=[question, max_turns, memory_strategy, temperature], outputs=[answer, trace], ) # .success() (not .then()) so the textbox is cleared ONLY on a clean run. # If Stop cancels the generator we leave the question intact, so the user # can tweak it and re-run without retyping. run_event.success( fn=lambda: "", inputs=[], outputs=[question], ) for btn, ex in zip(example_buttons, EXAMPLES): btn.click( fn=(lambda text=ex["text"]: text), inputs=[], outputs=[question], ) stop_btn.click(fn=None, cancels=[run_event]) clear_btn.click( fn=lambda: ("", "", "{}"), inputs=[], outputs=[question, answer, trace], ) # The research agent is almost entirely I/O-bound (it waits on the OSC vLLM # endpoint, Serper, Jina and Azure over HTTP), so many runs can proceed in # parallel on even a small CPU box. Gradio's default_concurrency_limit is 1, # which serialises every run_ui call and is what produces the "long queue of # requests pending" warning. Lift it; the real ceiling is the 8 OSC vLLM # instances behind the nginx load balancer. Tunable via the QUEST_CONCURRENCY # Space variable without a code change. QUEST_CONCURRENCY = int(os.getenv("QUEST_CONCURRENCY", "12")) QUEST_QUEUE_MAX = int(os.getenv("QUEST_QUEUE_MAX", "80")) demo.queue( default_concurrency_limit=QUEST_CONCURRENCY, max_size=QUEST_QUEUE_MAX, ) if __name__ == "__main__": demo.launch(max_threads=max(40, QUEST_CONCURRENCY * 3))