125 lines
3.9 KiB
Python
125 lines
3.9 KiB
Python
import requests
|
|
from typing import Any, Callable
|
|
|
|
import numpy as np
|
|
from chonkie.embeddings import BaseEmbeddings
|
|
|
|
from src.config import settings
|
|
|
|
|
|
class OllamaEmbeddings(BaseEmbeddings):
|
|
"""Chonkie embeddings adapter for a local Ollama embedding model."""
|
|
|
|
def __init__(
|
|
self,
|
|
model: str,
|
|
base_url: str = settings.ollama_local_url,
|
|
timeout: float = 60.0,
|
|
truncate: bool = True,
|
|
keep_alive: str = "5m",
|
|
) -> None:
|
|
self.model = model
|
|
self.base_url = base_url.rstrip("/")
|
|
self.timeout = timeout
|
|
self.truncate = truncate
|
|
self.keep_alive = keep_alive
|
|
self._dimension: int | None = None
|
|
|
|
@property
|
|
def dimension(self) -> int:
|
|
if self._dimension is None:
|
|
# Lazy-load the dimension from a real embedding response.
|
|
self._dimension = int(self.embed(" ").shape[0])
|
|
return self._dimension
|
|
|
|
def embed(self, text: str) -> np.ndarray:
|
|
embeddings = self._embed_api(text)
|
|
vector = np.asarray(embeddings[0], dtype=np.float32)
|
|
|
|
if self._dimension is None:
|
|
self._dimension = int(vector.shape[0])
|
|
|
|
return vector
|
|
|
|
def embed_batch(self, texts: list[str]) -> list[np.ndarray]:
|
|
if not texts:
|
|
return []
|
|
|
|
embeddings = self._embed_api(texts)
|
|
vectors = [np.asarray(vector, dtype=np.float32) for vector in embeddings]
|
|
|
|
if vectors and self._dimension is None:
|
|
self._dimension = int(vectors[0].shape[0])
|
|
|
|
return vectors
|
|
|
|
def count_tokens(self, text: str) -> int:
|
|
payload = self._build_payload(text)
|
|
response = self._post_embed(payload)
|
|
return int(response["prompt_eval_count"])
|
|
|
|
def count_tokens_batch(self, texts: list[str]) -> list[int]:
|
|
# Ollama returns a single prompt_eval_count for the whole request,
|
|
# not one count per input item, so we compute them individually.
|
|
return [self.count_tokens(text) for text in texts]
|
|
|
|
def get_tokenizer(self) -> Callable[[str], int]:
|
|
# Chonkie mainly needs something usable for token counting.
|
|
return self.count_tokens
|
|
|
|
@classmethod
|
|
def is_available(cls) -> bool:
|
|
try:
|
|
response = requests.get(
|
|
f"{settings.ollama_local_url}/api/tags",
|
|
timeout=5.0,
|
|
)
|
|
response.raise_for_status()
|
|
return True
|
|
except requests.RequestException:
|
|
return False
|
|
|
|
def __repr__(self) -> str:
|
|
return (
|
|
f"OllamaEmbeddings("
|
|
f"model={self.model!r}, "
|
|
f"base_url={self.base_url!r}, "
|
|
f"dimension={self._dimension!r}"
|
|
f")"
|
|
)
|
|
|
|
def _build_payload(self, text_or_texts: str | list[str]) -> dict[str, Any]:
|
|
return {
|
|
"model": self.model,
|
|
"input": text_or_texts,
|
|
"truncate": self.truncate,
|
|
"keep_alive": self.keep_alive,
|
|
}
|
|
|
|
def _post_embed(self, payload: dict[str, Any]) -> dict[str, Any]:
|
|
try:
|
|
response = requests.post(
|
|
f"{self.base_url}/api/embed",
|
|
json=payload,
|
|
timeout=self.timeout,
|
|
)
|
|
response.raise_for_status()
|
|
data = response.json()
|
|
except requests.RequestException as exc:
|
|
raise RuntimeError(
|
|
f"Failed to call Ollama embeddings endpoint at "
|
|
f"{self.base_url}/api/embed"
|
|
) from exc
|
|
|
|
if "embeddings" not in data:
|
|
raise RuntimeError(
|
|
"Ollama response did not include 'embeddings'. "
|
|
f"Response keys: {list(data.keys())}"
|
|
)
|
|
|
|
return data
|
|
|
|
def _embed_api(self, text_or_texts: str | list[str]) -> list[list[float]]:
|
|
payload = self._build_payload(text_or_texts)
|
|
data = self._post_embed(payload)
|
|
return data["embeddings"] |