File size: 10,178 Bytes
67cadad
a918698
67cadad
 
 
 
53aefb2
67cadad
 
 
 
a918698
 
 
 
 
 
 
67cadad
a918698
 
 
 
 
 
 
 
67cadad
 
 
53aefb2
 
 
67cadad
 
53aefb2
 
67cadad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
53aefb2
 
 
 
 
 
 
 
67cadad
 
 
 
a918698
 
 
 
 
 
 
 
 
 
 
 
67cadad
 
 
 
 
 
 
 
 
 
 
53aefb2
67cadad
53aefb2
67cadad
a918698
67cadad
 
 
53aefb2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a918698
 
53aefb2
 
 
 
 
 
 
 
a918698
 
53aefb2
a918698
 
 
 
53aefb2
 
 
 
 
 
a918698
 
 
 
 
 
 
 
 
53aefb2
0aaf5f7
53aefb2
 
 
a918698
53aefb2
 
 
a918698
53aefb2
a918698
 
 
53aefb2
 
 
 
 
 
 
 
a918698
 
 
 
 
 
 
 
 
53aefb2
 
67cadad
 
53aefb2
67cadad
53aefb2
 
 
67cadad
 
 
 
53aefb2
 
 
67cadad
 
53aefb2
 
 
 
 
67cadad
 
 
a918698
 
67cadad
53aefb2
 
 
a918698
 
 
 
 
 
67cadad
a918698
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67cadad
 
 
 
 
 
 
 
 
 
a918698
 
67cadad
 
a918698
 
 
 
 
 
 
 
 
 
 
67cadad
a918698
 
67cadad
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
"""LLM provider client for structured-output annotation.

Supports three OpenAI-compatible providers via parametric base URL:
- openrouter : https://openrouter.ai/api/v1   (MoE: many models behind one key)
- mistral    : https://api.mistral.ai/v1
- openai     : https://api.openai.com/v1
- ilaas      : https://llm.ilaas.fr/v1 (documentation: https://www.ilaas.fr/services-inference/)

All three accept the same OpenAI Chat Completions request shape, including
`response_format` (json_schema strict on OpenAI; json_object on Mistral; varies
on OpenRouter — we auto-fall back if the strict mode is rejected).
"""
from __future__ import annotations

import asyncio
import json
import time
from dataclasses import dataclass
from typing import Optional

import httpx

from schemas import AnnotationSchema, to_json_schema, validate as schema_validate
from prompts import VALIDATION_RETRY

DEFAULT_TIMEOUT = 60.0


BASE_URLS = {
    "openrouter": "https://openrouter.ai/api/v1",
    "mistral": "https://api.mistral.ai/v1",
    "openai": "https://api.openai.com/v1",
    "ilaas": "https://llm.ilaas.fr/v1",
}

PROVIDERS = tuple(BASE_URLS.keys())

CURATED_MODELS_BY_PROVIDER: dict[str, list[str]] = {
    "openrouter": [
        "openai/gpt-oss-20b:free",
        "google/gemma-4-26b-a4b-it:free",
        "meta-llama/llama-3.3-70b-instruct:free",
        "qwen/qwen3-next-80b-a3b-instruct:free",
        "deepseek/deepseek-v4-flash:free",
        "mistralai/mistral-nemo",
        "mistralai/mistral-small-24b-instruct-2501",
        "mistralai/ministral-3b-2512",
    ],
    "mistral": [
        "mistral-small-2603",
        "mistral-large-2512",
        "ministral-8b-2512",
        "ministral-3b-2512",
    ],
    "openai": [
        "gpt-5-mini-2025-08-07",
        "gpt-5-nano-2025-08-07",
        "gpt-5-2025-08-07",
        "gpt-4o-mini-2024-07-18",
    ],
    "ilaas": [
        "gemma-4-31b",
        "gpt-oss-120b",
        "llama-3.1-8b",
        "llama-3.3-70b",
        "qwen-3.6-35b-instruct",
        "mistral-small-3.2-24b",
    ]
}

# Back-compat alias used by other modules
CURATED_MODELS = CURATED_MODELS_BY_PROVIDER["openrouter"]


@dataclass
class ModelResult:
    model: str
    ok: bool
    annotation: Optional[dict]
    latency_s: float
    error: str = ""
    raw: str = ""


def _build_headers(provider: str, api_key: str) -> dict:
    h = {
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json",
    }
    if provider == "openrouter":
        h["HTTP-Referer"] = "https://lrec2026-llm-annotator.local"
        h["X-Title"] = "LREC2026 LLM-as-Annotator"
    return h

class LLMClient:
    def __init__(self, provider: str, api_key: str, timeout: float = DEFAULT_TIMEOUT):
        if provider not in BASE_URLS:
            raise ValueError(f"Unknown provider {provider!r}; expected one of {tuple(BASE_URLS)}")
        self.provider = provider
        self.api_key = api_key
        self.base_url = BASE_URLS[provider]
        self.endpoint = self.base_url + "/chat/completions"
        self.headers = _build_headers(provider, api_key)
        self.timeout = timeout
        self._client: httpx.AsyncClient | None = None

    async def __aenter__(self):
        self._client = httpx.AsyncClient(
            timeout=self.timeout,
            limits=httpx.Limits(
                max_connections=20,
                max_keepalive_connections=10,
            ),
        )
        return self

    async def __aexit__(self, exc_type, exc, tb):
        if self._client:
            await self._client.aclose()
            self._client = None

    async def annotate_one(
            self,
            *,
            system: str,
            user: str,
            schema: AnnotationSchema,
            model: str,
            temperature: float = 0.0,
            timeout: float = DEFAULT_TIMEOUT,
    ) -> ModelResult:
        """Call one model, validate JSON. One retry on schema-validation failure."""
        print(f"[LLM] start provider={self.provider} model={model}")
        json_schema = to_json_schema(schema)
        start = time.time()
        msgs = [{"role": "system", "content": system}, {"role": "user", "content": user}]
        try:
            client = self._client
            close_after = False
            if client is None:
                client = httpx.AsyncClient(timeout=timeout)
                close_after = True
            try:
                raw_text = await self._call(client, msgs, json_schema, model, temperature)
                ann, err = self._parse_and_validate(raw_text, schema)
                if err:
                    retry_msg = VALIDATION_RETRY + f"\n\nValidator errors:\n{err}\n\nPrevious response:\n{raw_text}"
                    msgs.append({"role": "assistant", "content": raw_text})
                    msgs.append({"role": "user", "content": retry_msg})
                    raw_text = await self._call(client, msgs, json_schema, model, temperature)
                    ann, err = self._parse_and_validate(raw_text, schema)
                if err:
                    print(
                        f"[LLM] error provider={self.provider} model={model} latency={time.time() - start:.2f}s error={err}")
                    return ModelResult(model=model, ok=False, annotation=None, latency_s=time.time() - start, error=err,
                                       raw=raw_text)
                print(f"[LLM] done provider={self.provider} model={model} latency={time.time() - start:.2f}s")
                return ModelResult(model=model, ok=True, annotation=ann, latency_s=time.time() - start, raw=raw_text)
            finally:
                if close_after:
                    await client.aclose()
        except Exception as e:
            print(f"[LLM] error provider={self.provider} model={model} latency={time.time() - start:.2f}s error={e}")
            return ModelResult(model=model, ok=False, annotation=None, latency_s=time.time() - start, error=str(e))

    async def annotate_many(
            self,
            *,
            models: list[str],
            system: str,
            user: str,
            schema: AnnotationSchema,
            temperature: float = 0.0,
            timeout: float = DEFAULT_TIMEOUT,
    ) -> list[ModelResult]:
        coros = [
            self.annotate_one(
                system=system, user=user, schema=schema, model=m, temperature=temperature, timeout=timeout
            )
            for m in models
        ]
        return await asyncio.gather(*coros)

    async def _call(self, client: httpx.AsyncClient, msgs: list[dict], json_schema: dict, model: str,
                    temperature: float) -> str:
        # Strict json_schema works on OpenAI and most OpenRouter models. For Mistral and
        # for some open-source models routed via OpenRouter, fall back to json_object.
        if self.provider in {"mistral", "ilaas"}:
            payload = {
                "model": model,
                "messages": msgs,
                "temperature": temperature,
                "response_format": {"type": "json_object"},
            }
        else:
            payload = {
                "model": model,
                "messages": msgs,
                "temperature": temperature,
                "response_format": {
                    "type": "json_schema",
                    "json_schema": {
                        "name": "annotation",
                        "strict": True,
                        "schema": json_schema,
                    },
                },
            }
        resp = await client.post(self.endpoint, headers=self.headers, json=payload)
        if resp.status_code >= 400:
            payload["response_format"] = {"type": "json_object"}
            resp = await client.post(self.endpoint, headers=self.headers, json=payload)
        if resp.status_code >= 400:
            payload.pop("response_format", None)
            resp = await client.post(self.endpoint, headers=self.headers, json=payload)
        resp.raise_for_status()
        data = resp.json()
        return data["choices"][0]["message"]["content"] or ""

    @staticmethod
    def _parse_and_validate(raw_text: str, schema: AnnotationSchema) -> tuple[Optional[dict], str]:
        text = (raw_text or "").strip()
        if text.startswith("```"):
            text = text.strip("`")
            if text.lower().startswith("json"):
                text = text[4:]
            text = text.strip()
        try:
            ann = json.loads(text)
        except json.JSONDecodeError as e:
            return None, f"Invalid JSON: {e}"
        ok, errors = schema_validate(schema, ann)
        if not ok:
            return None, "; ".join(errors[:10])
        return ann, ""


def test_connection_sync(api_key: str, provider: str = "openrouter", model: Optional[str] = None) -> tuple[bool, str]:
    """Quick blocking test for the 'Test' button in the key modal."""
    if provider not in BASE_URLS:
        return False, f"Unknown provider {provider!r}"
    if not model:
        models = CURATED_MODELS_BY_PROVIDER.get(provider) or []
        model = models[0] if models else None
    if not model:
        return False, "No model configured for this provider."
    url = BASE_URLS[provider] + "/chat/completions"
    try:
        resp = httpx.post(
            url,
            headers=_build_headers(provider, api_key),
            json={
                "model": model,
                "messages": [{"role": "user", "content": "Reply with the single word: OK"}],
                "max_tokens": 5,
                "temperature": 0.0,
            },
            timeout=20.0,
        )
        if resp.status_code >= 400:
            return False, f"HTTP {resp.status_code}: {resp.text[:200]}"
        content = resp.json()["choices"][0]["message"]["content"]
        return True, f"Connected ({provider} / {model}). Reply: {content!r}"
    except Exception as e:
        return False, str(e)


# Back-compat shim — old callers used OpenRouterClient(api_key=...).
class OpenRouterClient(LLMClient):
    def __init__(self, api_key: str, **kwargs):
        super().__init__(provider="openrouter", api_key=api_key)