File size: 15,736 Bytes
7eacaee 193b256 7eacaee | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 | """
SLM-Native Backends — First-class support for Small Language Models.
Purpose Agent is the world's first agentic framework designed natively for SLMs.
These backends handle the unique challenges of small models:
- Grammar-constrained JSON output (SLMs can't reliably produce JSON from prompts alone)
- Prompt compression for small context windows (8K-32K)
- Adaptive prompting (shorter system prompts, schema-first format)
- Token budget management
Supported backends:
- OllamaBackend: Local serving via Ollama (CPU/GPU, any GGUF model)
- LlamaCppBackend: Direct llama-cpp-python (CPU/Apple Silicon, GGUF)
- TransformersBackend: HuggingFace transformers (GPU, native weights)
All backends implement the same LLMBackend interface — swap freely.
"""
from __future__ import annotations
import json
import logging
import os
import re
from typing import Any, AsyncIterator, Iterator
from purpose_agent.llm_backend import ChatMessage, LLMBackend
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# SLM Prompt Compressor — reduces prompt size for small context windows
# ---------------------------------------------------------------------------
class SLMPromptCompressor:
"""
Compresses prompts for small context windows without losing critical info.
Strategies (from TinyAgent arxiv:2409.00608 + LLMLingua-2 arxiv:2403.12968):
1. Schema-first: Move JSON schema to top, compress descriptions
2. History truncation: Summarize old steps, keep recent ones verbatim
3. Example reduction: Fewer few-shot examples for SLMs
4. Whitespace stripping: Remove unnecessary formatting
No external dependencies — pure Python compression.
For better compression, install llmlingua: pip install llmlingua
"""
def __init__(self, max_tokens: int = 4096, aggressive: bool = False):
self.max_tokens = max_tokens
self.aggressive = aggressive
def compress(self, text: str, budget: int | None = None) -> str:
"""Compress text to fit within token budget."""
budget = budget or self.max_tokens
# Rough estimate: 1 token ≈ 4 chars
char_budget = budget * 4
if len(text) <= char_budget:
return text
compressed = text
# Stage 1: Strip excessive whitespace
compressed = re.sub(r'\n{3,}', '\n\n', compressed)
compressed = re.sub(r'[ \t]{2,}', ' ', compressed)
compressed = re.sub(r'^\s+', '', compressed, flags=re.MULTILINE)
if len(compressed) <= char_budget:
return compressed
# Stage 2: Shorten verbose sections
if self.aggressive:
# Remove markdown formatting
compressed = re.sub(r'\*\*([^*]+)\*\*', r'\1', compressed)
compressed = re.sub(r'#{1,3}\s+', '', compressed)
# Shorten common verbose phrases
replacements = {
"You MUST respond with": "Respond with",
"Based on the current state and your goal, ": "",
"Respond in this exact JSON format:": "JSON format:",
"Step-by-step justification": "Justification",
"Specific observable state changes": "State changes",
}
for old, new in replacements.items():
compressed = compressed.replace(old, new)
if len(compressed) <= char_budget:
return compressed
# Stage 3: Truncate from middle (keep start + end)
keep_start = char_budget * 2 // 3
keep_end = char_budget // 3
compressed = compressed[:keep_start] + "\n...[truncated]...\n" + compressed[-keep_end:]
return compressed
def compress_messages(
self, messages: list[ChatMessage], budget: int | None = None
) -> list[ChatMessage]:
"""Compress a message list to fit within token budget."""
budget = budget or self.max_tokens
total_chars = sum(len(m.content) for m in messages)
char_budget = budget * 4
if total_chars <= char_budget:
return messages
result = []
# Always keep system prompt (compress it), always keep last user message
for i, msg in enumerate(messages):
if msg.role == "system":
result.append(ChatMessage(
role="system",
content=self.compress(msg.content, budget=budget // 3),
))
elif i == len(messages) - 1:
# Last message — keep more of it
result.append(ChatMessage(
role=msg.role,
content=self.compress(msg.content, budget=budget // 2),
))
else:
result.append(ChatMessage(
role=msg.role,
content=self.compress(msg.content, budget=budget // 4),
))
return result
# ---------------------------------------------------------------------------
# Ollama Backend — Best for local SLMs
# ---------------------------------------------------------------------------
class OllamaBackend(LLMBackend):
"""
Local model serving via Ollama with grammar-constrained JSON output.
Ollama's grammar engine (via llama.cpp) forces valid JSON output from
ANY model — even tiny ones that can't produce reliable JSON from prompts.
This is the key advantage for SLM agent use.
Setup:
1. Install Ollama: https://ollama.ai
2. Pull a model: ollama pull qwen3:1.7b
3. Use this backend:
Example:
backend = OllamaBackend(model="qwen3:1.7b") # 1.7B params, runs on CPU
backend = OllamaBackend(model="llama3.2:1b") # 1B params, ultra-light
backend = OllamaBackend(model="phi4-mini") # 3.8B, best tool-use
backend = OllamaBackend(model="smollm2:1.7b") # HF native SLM
Also works with large models:
backend = OllamaBackend(model="qwen3:32b") # Full LLM
"""
def __init__(
self,
model: str = "qwen3:1.7b",
host: str = "http://localhost:11434",
context_window: int = 8192,
compress_prompts: bool = True,
num_ctx: int | None = None,
):
self.model = model
self.host = host
self.context_window = context_window
self.compress_prompts = compress_prompts
self.num_ctx = num_ctx or context_window
self.compressor = SLMPromptCompressor(
max_tokens=context_window, aggressive=(context_window <= 8192)
)
self._token_count = 0
def _get_client(self):
"""Lazy import ollama client."""
try:
from ollama import Client
return Client(host=self.host)
except ImportError:
raise ImportError(
"Ollama client not installed. Run: pip install ollama\n"
"Also install Ollama server: https://ollama.ai"
)
def generate(
self,
messages: list[ChatMessage],
temperature: float = 0.7,
max_tokens: int = 2048,
stop: list[str] | None = None,
) -> str:
client = self._get_client()
if self.compress_prompts:
messages = self.compressor.compress_messages(messages, self.context_window)
msg_dicts = [{"role": m.role, "content": m.content} for m in messages]
response = client.chat(
model=self.model,
messages=msg_dicts,
options={
"temperature": temperature,
"num_predict": max_tokens,
"num_ctx": self.num_ctx,
"stop": stop or [],
},
)
content = self._strip_thinking(response.message.content or "")
self._token_count += response.get("eval_count", 0) + response.get("prompt_eval_count", 0)
return content
def generate_structured(
self,
messages: list[ChatMessage],
schema: dict[str, Any],
temperature: float = 0.3,
max_tokens: int = 1024,
) -> dict[str, Any]:
"""
Grammar-constrained JSON generation.
Ollama uses llama.cpp's grammar engine to FORCE valid JSON output
matching the schema. This works even with tiny models that can't
produce valid JSON from prompts alone.
"""
client = self._get_client()
if self.compress_prompts:
messages = self.compressor.compress_messages(messages, self.context_window)
msg_dicts = [{"role": m.role, "content": m.content} for m in messages]
response = client.chat(
model=self.model,
messages=msg_dicts,
format=schema, # Grammar-constrained output!
options={
"temperature": temperature,
"num_predict": max_tokens,
"num_ctx": self.num_ctx,
},
)
content = response.message.content or "{}"
self._token_count += response.get("eval_count", 0) + response.get("prompt_eval_count", 0)
return json.loads(content)
def generate_stream(
self,
messages: list[ChatMessage],
temperature: float = 0.7,
max_tokens: int = 2048,
) -> Iterator[str]:
"""Streaming generation — yields tokens as they're produced."""
client = self._get_client()
if self.compress_prompts:
messages = self.compressor.compress_messages(messages, self.context_window)
msg_dicts = [{"role": m.role, "content": m.content} for m in messages]
stream = client.chat(
model=self.model,
messages=msg_dicts,
stream=True,
options={
"temperature": temperature,
"num_predict": max_tokens,
"num_ctx": self.num_ctx,
},
)
for chunk in stream:
token = chunk.get("message", {}).get("content", "")
if token:
yield token
@property
def total_tokens(self) -> int:
return self._token_count
# ---------------------------------------------------------------------------
# LlamaCpp Backend — Direct CPU/Apple Silicon/GGUF
# ---------------------------------------------------------------------------
class LlamaCppBackend(LLMBackend):
"""
Direct llama-cpp-python backend for GGUF models.
Best for: CPU inference, Apple Silicon, edge deployment, offline use.
Example:
backend = LlamaCppBackend(model_path="./qwen2.5-1.5b-instruct-q4_k_m.gguf")
backend = LlamaCppBackend(
model_path="./phi-4-mini-q4.gguf",
n_ctx=4096,
n_gpu_layers=35, # Offload to GPU
)
"""
def __init__(
self,
model_path: str,
n_ctx: int = 4096,
n_gpu_layers: int = 0,
verbose: bool = False,
):
try:
from llama_cpp import Llama
except ImportError:
raise ImportError("llama-cpp-python not installed. Run: pip install llama-cpp-python")
self.model_path = model_path
self.llm = Llama(
model_path=model_path,
n_ctx=n_ctx,
n_gpu_layers=n_gpu_layers,
verbose=verbose,
)
self.compressor = SLMPromptCompressor(max_tokens=n_ctx, aggressive=True)
self._token_count = 0
def generate(
self,
messages: list[ChatMessage],
temperature: float = 0.7,
max_tokens: int = 2048,
stop: list[str] | None = None,
) -> str:
messages = self.compressor.compress_messages(messages)
msg_dicts = [{"role": m.role, "content": m.content} for m in messages]
response = self.llm.create_chat_completion(
messages=msg_dicts,
temperature=temperature,
max_tokens=max_tokens,
stop=stop,
)
content = response["choices"][0]["message"]["content"] or ""
usage = response.get("usage", {})
self._token_count += usage.get("total_tokens", 0)
return content
def generate_structured(
self,
messages: list[ChatMessage],
schema: dict[str, Any],
temperature: float = 0.3,
max_tokens: int = 1024,
) -> dict[str, Any]:
"""Grammar-constrained JSON via llama.cpp GBNF grammar."""
from llama_cpp import LlamaGrammar
grammar = LlamaGrammar.from_json_schema(json.dumps(schema))
messages = self.compressor.compress_messages(messages)
msg_dicts = [{"role": m.role, "content": m.content} for m in messages]
response = self.llm.create_chat_completion(
messages=msg_dicts,
temperature=temperature,
max_tokens=max_tokens,
grammar=grammar,
)
content = response["choices"][0]["message"]["content"] or "{}"
usage = response.get("usage", {})
self._token_count += usage.get("total_tokens", 0)
return json.loads(content)
def generate_stream(
self,
messages: list[ChatMessage],
temperature: float = 0.7,
max_tokens: int = 2048,
) -> Iterator[str]:
messages = self.compressor.compress_messages(messages)
msg_dicts = [{"role": m.role, "content": m.content} for m in messages]
stream = self.llm.create_chat_completion(
messages=msg_dicts,
temperature=temperature,
max_tokens=max_tokens,
stream=True,
)
for chunk in stream:
delta = chunk.get("choices", [{}])[0].get("delta", {})
token = delta.get("content", "")
if token:
yield token
@property
def total_tokens(self) -> int:
return self._token_count
# ---------------------------------------------------------------------------
# Model Registry — Easy model selection for SLMs
# ---------------------------------------------------------------------------
# Recommended SLMs for agent tasks, ranked by capability
SLM_REGISTRY = {
# Model ID → (Ollama name, context window, description)
"phi-4-mini": ("phi4-mini", 16384, "3.8B, best schema compliance, Microsoft"),
"qwen3-1.7b": ("qwen3:1.7b", 32768, "1.7B, strong function calling, 32K context"),
"qwen3-0.6b": ("qwen3:0.6b", 32768, "0.6B, ultra-light, 32K context"),
"qwen2.5-1.5b": ("qwen2.5:1.5b", 32768, "1.5B, proven tool-use"),
"llama-3.2-3b": ("llama3.2:3b", 131072, "3B, 128K context, Meta"),
"llama-3.2-1b": ("llama3.2:1b", 131072, "1B, smallest Llama, 128K context"),
"smollm2-1.7b": ("smollm2:1.7b", 8192, "1.7B, HF native, 8K context (tight!)"),
"gemma-3-1b": ("gemma3:1b", 32768, "1B, Google, multimodal capable"),
}
def create_slm_backend(
model_key: str = "qwen3-1.7b",
host: str = "http://localhost:11434",
) -> OllamaBackend:
"""
Create an SLM backend from the registry.
Usage:
backend = create_slm_backend("phi-4-mini") # Best overall
backend = create_slm_backend("qwen3-0.6b") # Ultra-light
backend = create_slm_backend("llama-3.2-1b") # Smallest Llama
"""
if model_key not in SLM_REGISTRY:
available = ", ".join(SLM_REGISTRY.keys())
raise ValueError(f"Unknown SLM '{model_key}'. Available: {available}")
ollama_name, ctx_window, desc = SLM_REGISTRY[model_key]
logger.info(f"Creating SLM backend: {model_key} ({desc})")
return OllamaBackend(
model=ollama_name,
host=host,
context_window=ctx_window,
compress_prompts=True,
)
|