andnp-searchkernel 0.4.1

Source-agnostic search/indexing kernel: domain-agnostic content ingestion, hybrid vector+keyword+graph search, pluggable embedding/LLM/reranker providers.

andnp-searchkernel

A domain-agnostic search/indexing kernel for building hybrid vector + keyword + graph search systems with pluggable embedding, LLM, and reranker providers.

Status

Pre-alpha, canonical record path established. The legacy chunk query pipeline has been removed; source adapters, backend integrations, and performance validation remain subject to change.

Optional backends

The core package provides the domain models, ports, record search pipeline, and evaluation primitives. Install only the integrations required by an application:

pip install andnp-searchkernel[pgvector,huggingface,markdown]

Available extras are faiss, pgvector, huggingface, and markdown. FAISS and pgvector implement the same record-oriented backend contracts; they can be selected independently or used together during migrations.

Canonical search composition

Compose local search from the record-oriented ports:

from searchkernel.api import SearchKernel

kernel = SearchKernel.build(
    record_hydrator=record_hydrator,
    keyword_store=keyword_store,
    vector_store=vector_store,
    graph_store=graph_store,
    embedding_provider=embedding_provider,
)

SearchKernel.build registers a canonical SearchOrchestrator for these dependencies. Callers that already own one may pass orchestrator= instead. The deprecated chunk-oriented execution path has been removed; migrate callers to this record composition. For an explicit migration bridge around an existing legacy orchestrator, import LegacyLocalOrchestratorAdapter from searchkernel.runtime.local; it is not the supported search pipeline.

Integration tests

The pgvector integration tests automatically start a temporary pgvector/pgvector:pg17 Docker container when SEARCHKERNEL_PG_DSN is not set. Docker must be running:

uv run pytest tests/integration

To use an existing PostgreSQL instance instead, set SEARCHKERNEL_PG_DSN to its connection string. The database must allow the vector extension to be created.

Releases

Merges to main with feat, fix, or breaking Conventional Commits are released automatically. The release workflow bumps the SemVer version, updates pyproject.toml and uv.lock, pushes a v* tag, and dispatches the PyPI publishing workflow. Documentation, chore, and test-only commits do not create releases.

License

MIT License. See LICENSE for details.