Milvus Lite pilot

Milvus Lite is an opt-in pilot and is not wired into rag.py, the router, or the normal LanceDB ingestion path. The repository therefore retains its offline, no-service default behavior.

Install and use

Milvus is an optional package extra:

uv sync --extra milvus-lite

The adapter is hpe_networking_mcp.pipeline.clients.milvus_client. It opens only a local .db path, creating parent directories as needed:

from hpe_networking_mcp.pipeline.clients.milvus_client import MilvusLiteStore

store = MilvusLiteStore("data/milvus-lite.db")
store.upsert([{
    "id": "chunk-1",
    "text": "WPA3 uses SAE",
    "source": "developer_docs",
    "vector": [0.1, 0.2],
}])
hits = store.search([0.1, 0.2], metadata_filter={"source": "developer_docs"})

HPE_MCP_MILVUS_PATH supplies the default path when the constructor argument is omitted. IDs are stable across re-embedding: supplied IDs are preserved, otherwise a SHA-256 ID is derived from content/provenance fields.

Dense retrieval and bounded scalar equality/in metadata filters are tested without Milvus installed. Hybrid retrieval is capability-detected. Because PyMilvus hybrid request/ranker constructors vary and Milvus Lite support can differ by release, the pilot accepts installed-client-native request objects and raises a clear MilvusCapabilityError when hybrid_search is unavailable or incompatible. It never silently falls back to a remote service.

If the optional package is absent, construction raises MilvusDependencyError with the install command; importing the module itself continues to work.

Corpus comparison

Build a local pilot from the existing LanceDB vectors and compare dense Milvus retrieval on the active how-to evaluation questions:

uv sync --extra milvus-lite
uv run python scripts/benchmark_milvus.py \
  --milvus-path data/milvus-lite.db \
  --rebuild \
  --json milvus-benchmark.json

The report separates the one-time copy cost from warm search p50/p95 and records source-hit, keyword, and MRR results. Compare it with scripts/benchmark_rag.py and tests/eval/run_eval.py; do not promote Milvus to the default unless it preserves the full hybrid quality gate, not merely dense-search latency.

Current local result

On the 262,104-row migrated corpus, the dense-only pilot copied the corpus in about 94 seconds and measured 283.7 ms warm p50 / 331.6 ms p95 across the 16 active how-to questions, with 0.812 source-hit@5, 0.750 keyword-hit, and 0.740 MRR. The LanceDB baseline remains 0.972 source-hit@5, 0.938 how-to-recall@5, and 0.972 MRR on the full evaluation gate, with roughly 10 ms warm p50 on the representative local benchmark. Milvus Lite therefore remains an opt-in experiment; its dense-only result is neither faster nor accurate enough to replace LanceDB.