122 lines
4.5 KiB
Python
122 lines
4.5 KiB
Python
"""Adapter Ollama — implémentation concrète des ports LLMProvider et LLMChatProvider.
|
|
|
|
Isole le reste de l'application des spécificités du protocole Ollama
|
|
(URL /api/generate, /api/chat, payload, parsing). Pour swap vers OpenAI
|
|
demain, on écrit un nouvel adapter sans toucher au reste du code.
|
|
"""
|
|
import json
|
|
from typing import AsyncIterator
|
|
|
|
import httpx
|
|
|
|
from app.core.config import Settings
|
|
from app.domain.models import ChatMessage
|
|
from app.domain.ports import LLMProviderError
|
|
|
|
|
|
class OllamaLLMProvider:
|
|
"""Implémentation des ports LLM — appelle un serveur Ollama via HTTP.
|
|
|
|
Satisfait implicitement (duck typing) à la fois `LLMProvider` (endpoint
|
|
/api/generate, appel unique) et `LLMChatProvider` (endpoint /api/chat,
|
|
streaming token par token).
|
|
"""
|
|
|
|
def __init__(self, settings: Settings) -> None:
|
|
self._base_url = settings.ollama_base_url
|
|
self._model = settings.llm_model
|
|
self._timeout = settings.llm_timeout_seconds
|
|
self._num_ctx = settings.llm_num_ctx
|
|
|
|
def _build_options(self, temperature: float | None) -> dict[str, object]:
|
|
"""Construit le dict `options` attendu par Ollama (hyperparamètres).
|
|
|
|
`num_ctx` est TOUJOURS envoyé — sinon Ollama retombe sur son défaut
|
|
2048 et tronque silencieusement les gros prompts (Structural Context
|
|
du Lore enrichi depuis b9). `temperature` n'est ajoutée que si
|
|
fournie par le use case (sinon Ollama utilise son défaut).
|
|
"""
|
|
options: dict[str, object] = {"num_ctx": self._num_ctx}
|
|
if temperature is not None:
|
|
options["temperature"] = temperature
|
|
return options
|
|
|
|
async def generate(
|
|
self,
|
|
prompt: str,
|
|
*,
|
|
output_format: str | None = None,
|
|
temperature: float | None = None,
|
|
) -> str:
|
|
url = f"{self._base_url}/api/generate"
|
|
payload: dict[str, object] = {
|
|
"model": self._model,
|
|
"prompt": prompt,
|
|
"stream": False,
|
|
"options": self._build_options(temperature),
|
|
}
|
|
if output_format is not None:
|
|
payload["format"] = output_format
|
|
|
|
async with httpx.AsyncClient(timeout=self._timeout) as client:
|
|
try:
|
|
response = await client.post(url, json=payload)
|
|
response.raise_for_status()
|
|
except httpx.HTTPError as exc:
|
|
raise LLMProviderError(
|
|
f"Erreur lors de l'appel à Ollama : {exc}"
|
|
) from exc
|
|
|
|
return response.json()["response"]
|
|
|
|
async def stream_chat(
|
|
self,
|
|
messages: list[ChatMessage],
|
|
*,
|
|
system_prompt: str | None = None,
|
|
temperature: float | None = None,
|
|
) -> AsyncIterator[str]:
|
|
"""Streame depuis Ollama /api/chat. Parse le NDJSON ligne par ligne.
|
|
|
|
Ollama renvoie un JSON par ligne au fil de la génération :
|
|
- étapes intermédiaires : `{"message": {"content": "token"}, "done": false}`
|
|
- étape finale : `{"done": true, ...}`
|
|
|
|
On yield chaque token non-vide au consommateur, qui se charge du
|
|
formatage SSE (c'est la responsabilité du controller HTTP, pas
|
|
de l'adapter LLM).
|
|
"""
|
|
url = f"{self._base_url}/api/chat"
|
|
|
|
payload_messages: list[dict[str, str]] = []
|
|
if system_prompt:
|
|
payload_messages.append({"role": "system", "content": system_prompt})
|
|
payload_messages.extend(
|
|
{"role": m.role, "content": m.content} for m in messages
|
|
)
|
|
|
|
payload: dict[str, object] = {
|
|
"model": self._model,
|
|
"messages": payload_messages,
|
|
"stream": True,
|
|
"options": self._build_options(temperature),
|
|
}
|
|
|
|
async with httpx.AsyncClient(timeout=self._timeout) as client:
|
|
try:
|
|
async with client.stream("POST", url, json=payload) as response:
|
|
response.raise_for_status()
|
|
async for line in response.aiter_lines():
|
|
if not line.strip():
|
|
continue
|
|
chunk = json.loads(line)
|
|
if chunk.get("done"):
|
|
break
|
|
token = chunk.get("message", {}).get("content", "")
|
|
if token:
|
|
yield token
|
|
except httpx.HTTPError as exc:
|
|
raise LLMProviderError(
|
|
f"Erreur lors du streaming Ollama : {exc}"
|
|
) from exc
|