mistralai-search-toolkit-plugins-qdrant

Qdrant store for mistralai-search-toolkit.

Same three methods as the other backends: index, search, delete. Navigation and patch work too if you want them.

Install

uv add mistralai-search-toolkit-plugins-qdrant
docker run -p 6333:6333 qdrant/qdrant

Usage

from mistralai.search.toolkit.embedding import CustomEmbeddingModel
from mistralai.search.toolkit.plugins.qdrant import (
    QdrantApp,
    QdrantCollectionSchema,
    QdrantConnectionConfig,
)
from mistralai.search.toolkit.search import VectorSearchQuery

schema = QdrantCollectionSchema(
    collection_name="docs",
    document_type=MyDoc,
    embedding_model=CustomEmbeddingModel(name="my-embedder", dimensions=1024),
)
app = QdrantApp([schema])
config = QdrantConnectionConfig(url="http://localhost:6333")

await app.create_collection(config, "docs")
store = app.get_search_index(config, "docs")
try:
    await store.index_document(doc)
    results = await store.search(VectorSearchQuery(embedding=vec, top_k=10))
finally:
    await store.aclose()

If you already have an AsyncQdrantClient, pass that instead of a config. Tests can skip the server:

QdrantConnectionConfig(path=":memory:")

One Qdrant point per chunk. Reindex writes the new points first, then drops whatever is left from the old version, so a crash does not empty the document.

exclude_ids works. max_candidates becomes hnsw_ef.

Give both embedding and query and you get hybrid search: dense rank fused with full-text over content by weighted RRF. Tune it with QdrantSearchQuery. The fusion score only means anything inside that one result set. distance comes from the vector side. A hit that only the text side found has distance=None.

navigate and read use [start, end) offsets. grep matches tokens. PHRASE cares about order, TERM does not.

patch_chunk and patch_document merge metadata one key at a time. Set a key to None to drop it.

Extra fields on a chunk keep their name. Extra fields on a document get a document_ prefix.

from typing import Annotated
from mistralai.search.toolkit.plugins.qdrant import QdrantField

section: Annotated[str | None, QdrantField(name="section_title")] = None
internal: Annotated[str | None, QdrantField(ignore=True)] = None

Qdrant only accepts integer or UUID point ids. The usual UUID5 chunk ids go through as-is. Anything else is hashed with uuid5, and the original string stays in payload id.

uv sync --group dev
uv run pytest tests/ -q