57 lines
1.6 KiB
Python
57 lines
1.6 KiB
Python
import asyncio
|
|
import logging
|
|
from functools import lru_cache
|
|
|
|
from ollama import AsyncClient
|
|
|
|
from app.config import get_settings
|
|
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
_BACKOFF_SECONDS: tuple[int, ...] = (1, 2, 4)
|
|
|
|
|
|
class EmbeddingError(Exception):
|
|
pass
|
|
|
|
|
|
@lru_cache(maxsize=1)
|
|
def _client() -> AsyncClient:
|
|
return AsyncClient(host=get_settings().ollama_url)
|
|
|
|
|
|
async def embed_texts(texts: list[str], model: str) -> list[list[float]]:
|
|
"""Embed each text via Ollama. Retries individual calls 3x with backoff."""
|
|
vectors: list[list[float]] = []
|
|
for text in texts:
|
|
vec = await _embed_one(text, model)
|
|
vectors.append(vec)
|
|
return vectors
|
|
|
|
|
|
async def _embed_one(text: str, model: str) -> list[float]:
|
|
last_err: Exception | None = None
|
|
client = _client()
|
|
for attempt in range(len(_BACKOFF_SECONDS) + 1):
|
|
try:
|
|
response = await client.embeddings(model=model, prompt=text)
|
|
return list(response["embedding"])
|
|
except Exception as exc:
|
|
last_err = exc
|
|
if attempt < len(_BACKOFF_SECONDS):
|
|
wait = _BACKOFF_SECONDS[attempt]
|
|
logger.warning(
|
|
"ollama embed retry",
|
|
extra={"event": "embed_retry", "attempt": attempt + 1, "wait_s": wait, "error": str(exc)},
|
|
)
|
|
await asyncio.sleep(wait)
|
|
raise EmbeddingError("embed failed after retries") from last_err
|
|
|
|
|
|
async def embedding_dimension(model: str) -> int:
|
|
"""Probe a single embedding to discover the model's vector dimension."""
|
|
vec = await _embed_one("dimension probe", model)
|
|
return len(vec)
|