Files
breakpilot-core/rag-service/embedding_client.py
Benjamin Admin 92ca5b7ba5 feat(rag): use Ollama for embeddings instead of embedding-service
Switch to Ollama's bge-m3 model (1024-dim) for generating embeddings,
solving the dimension mismatch with Qdrant collections. Embedding-service
still used for chunking, reranking, and PDF extraction.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-27 07:46:57 +01:00

163 lines
5.6 KiB
Python

import logging
import os
from typing import Optional
import httpx
from config import settings
logger = logging.getLogger("rag-service.embedding")
_TIMEOUT = httpx.Timeout(timeout=120.0, connect=10.0)
_EMBED_TIMEOUT = httpx.Timeout(timeout=300.0, connect=10.0)
# Ollama config for embeddings (bge-m3, 1024-dim)
_OLLAMA_URL = os.getenv("OLLAMA_URL", "http://ollama:11434")
_OLLAMA_EMBED_MODEL = os.getenv("OLLAMA_EMBED_MODEL", "bge-m3")
# Batch size for Ollama embedding requests
_EMBED_BATCH_SIZE = int(os.getenv("EMBED_BATCH_SIZE", "32"))
class EmbeddingClient:
"""
Hybrid client:
- Embeddings via Ollama (bge-m3, 1024-dim) for Qdrant compatibility
- Chunking + PDF extraction via embedding-service (port 8087)
"""
def __init__(self) -> None:
self._embed_svc_url: str = settings.EMBEDDING_SERVICE_URL.rstrip("/")
self._ollama_url: str = _OLLAMA_URL.rstrip("/")
self._embed_model: str = _OLLAMA_EMBED_MODEL
def _svc_url(self, path: str) -> str:
return f"{self._embed_svc_url}{path}"
# ------------------------------------------------------------------
# Embeddings (via Ollama)
# ------------------------------------------------------------------
async def generate_embeddings(self, texts: list[str]) -> list[list[float]]:
"""
Generate embeddings via Ollama's bge-m3 model.
Processes in batches to avoid timeout on large uploads.
"""
all_embeddings: list[list[float]] = []
for i in range(0, len(texts), _EMBED_BATCH_SIZE):
batch = texts[i : i + _EMBED_BATCH_SIZE]
batch_embeddings = []
async with httpx.AsyncClient(timeout=_EMBED_TIMEOUT) as client:
for text in batch:
response = await client.post(
f"{self._ollama_url}/api/embeddings",
json={
"model": self._embed_model,
"prompt": text,
},
)
response.raise_for_status()
data = response.json()
embedding = data.get("embedding", [])
if not embedding:
raise ValueError(
f"Ollama returned empty embedding for model {self._embed_model}"
)
batch_embeddings.append(embedding)
all_embeddings.extend(batch_embeddings)
if i + _EMBED_BATCH_SIZE < len(texts):
logger.info(
"Embedding progress: %d/%d", len(all_embeddings), len(texts)
)
return all_embeddings
async def generate_single_embedding(self, text: str) -> list[float]:
"""Convenience wrapper for a single text."""
results = await self.generate_embeddings([text])
if not results:
raise ValueError("Ollama returned empty result")
return results[0]
# ------------------------------------------------------------------
# Reranking (via embedding-service)
# ------------------------------------------------------------------
async def rerank_documents(
self,
query: str,
documents: list[str],
top_k: int = 10,
) -> list[dict]:
"""
Ask the embedding service to re-rank documents for a given query.
Returns a list of {index, score, text}.
"""
async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
response = await client.post(
self._svc_url("/rerank"),
json={
"query": query,
"documents": documents,
"top_k": top_k,
},
)
response.raise_for_status()
data = response.json()
return data.get("results", [])
# ------------------------------------------------------------------
# Chunking (via embedding-service)
# ------------------------------------------------------------------
async def chunk_text(
self,
text: str,
strategy: str = "recursive",
chunk_size: int = 512,
overlap: int = 50,
) -> list[str]:
"""
Ask the embedding service to chunk a long text.
Returns a list of chunk strings.
"""
async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
response = await client.post(
self._svc_url("/chunk"),
json={
"text": text,
"strategy": strategy,
"chunk_size": chunk_size,
"overlap": overlap,
},
)
response.raise_for_status()
data = response.json()
return data.get("chunks", [])
# ------------------------------------------------------------------
# PDF extraction (via embedding-service)
# ------------------------------------------------------------------
async def extract_pdf(self, pdf_bytes: bytes) -> str:
"""
Send raw PDF bytes to the embedding service for text extraction.
Returns the extracted text as a string.
"""
async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
response = await client.post(
self._svc_url("/extract-pdf"),
files={"file": ("document.pdf", pdf_bytes, "application/pdf")},
)
response.raise_for_status()
data = response.json()
return data.get("text", "")
# Singleton
embedding_client = EmbeddingClient()