Hands-on Labs

English · 中文版

Install from the repository root:

uv sync

Every lab is deterministic and uses temporary storage where persistence is needed.

1. Payload filtering

uv run python -m miniqdrant.labs.filtering

Expected: Matching ids: [1, 3] and a printed selected plan. Point 2 is vector-similar but excluded because it belongs to tenant b. Watch the public collection API combine vector scoring with a payload-index candidate set.

2. Immutable segments and optimization

uv run python -m miniqdrant.labs.segments

Expected:

Segments before optimize: 2
Segments after optimize: 1

Each upsert-plus-flush publishes one immutable segment. optimize() compacts the two published segments into one while preserving collection contents.

3. Close and reopen recovery

uv run python -m miniqdrant.labs.recovery

Expected: Restored ids: [1, 2]. The lab flushes, closes, and opens a new Database, forcing metadata and segment recovery through the public lifecycle boundary.

4. Compare four search plans

uv run python -m miniqdrant.labs.plan_comparison

Expected plan labels are exact_full_scan, hnsw, filtered_hnsw, and quantized_hnsw_rescore, with a visited_count beside each. Compare work and IDs, but read Differences from Qdrant: the quantized branch is a decoded-int8 full scan plus float rescore despite its teaching plan name.

5. HNSW recall

uv run python -m miniqdrant.labs.recall

The fixed seed runs 5 queries over 80 points and prints mean recall@5 (currently 1.000 for this fixture). Exact search supplies the reference set; the metric counts how many exact top-five IDs also appear in HNSW candidates. This is a reproducible mechanism check, not a production benchmark.

Use the behavior matrix to continue with focused tests, or run all tests with uv run pytest -q.