Qdrant: version matching + multiple collections (verified 2026-09-04)

qdrant-client must match the server minor version

Symptom chain with client 1.19.0 vs server 1.17.1 (qdrant/qdrant:v1.17.1):

Qdrant client version 1.19.0 is incompatible with server version 1.17.1 warning,
create_collection with a bare models.VectorParams(size=1024, ...) silently creates an ANONYMOUS vector (name ""), NOT dense,
the subsequent upsert fails: 400 ... Not existing vector name error: dense.

Fix — pin the client to the server minor:

pip install "qdrant-client==1.17.1"     # match qdrant/qdrant:v1.17.1

Always pass NAMED vectors so the config works regardless of client version:

from qdrant_client import QdrantClient, models
c = QdrantClient("http://localhost:6333")
c.create_collection(
    collection_name=NAME,
    vectors_config={
        "dense": models.VectorParams(size=1024, distance=models.Distance.COSINE),
    },
    sparse_vectors_config={
        "sparse": models.SparseVectorParams(index=models.SparseIndexParams(on_disk=True)),
    },
)

Query API in qdrant-client 1.17

client.search(...) does NOT exist in 1.17.
query_points(..., query_vector=...) → AssertionError: Unknown arguments: ['query_vector'].
Working call:

res = c.query_points(
    collection_name=NAME,
    query=<dense_embedding_list>,   # list[float] from Ollama /api/embeddings
    using="dense",                  # named-vector selector
    limit=5,
    with_payload=True,
)
for pt in res.points:
    print(pt.score, pt.payload.get("text"))

One collection per project/domain (multi-collection design)

For a distinct document set (batch of PDF protocol/files), create a SEPARATE collection with the SAME schema (dense 1024d COSINE + sparse sparse BM25) and the SAME embedder (bge-m3). Keep search query embeddings compatible by using the same embedder for all collections. Benefits: independent re-index, per-domain context search, no pollution of the general KB. Example: skc_vinny_gorod alongside knowledge_base.

Do NOT retarget context_enhancer to a second collection

context_enhancer.py binds COLLECTION = os.environ.get("QDRANT_COLLECTION", "knowledge_base") at IMPORT time (module level). Swapping the env var at runtime does NOT retarget it — the module-level constant is already fixed.

To search a second collection, write a STANDALONE REST search:

POST http://ollama:11434/api/embeddings {"model": "bge-m3", "prompt": text} → embedding (1024d).
POST http://qdrant:6333/collections/<NAME>/points/query with {"vector": emb, "limit": N, "with_payload": true, "using": "dense"}.
Read result.points[*].payload + .score.

This same pattern is the foundation for a per-project context-injector hook (e.g. NetBox project context) — hit the secondary collection directly, don't go through the shared KB search.

Scanned PDFs: pymupdf returns empty text (no text layer)

page.get_text("text") returns "" for image-only pages — verified on a real 51-page government PDF (0 text on every page). It is a genuine scan, not a glitch. Mark scanned pages [SCANNED_PAGE] and route to OCR (marker-pdf / vision); do NOT report "no content" or fabricate text. pymupdf's built-in Type1 fonts (times-roman, helv, cour, tiro) do NOT contain the Cyrillic glyph map — inserting Cyrillic with them renders as dots and re-extracts as dots. Real PDFs (generated from Word/CAD) embed proper fonts and extract Cyrillic fine; the byte test above is only a pymupdf-font artifact, not a real-PDF problem.

3.5 KiB Raw Permalink Blame History