andnp-searchkernel

andnp-searchkernel is a source-agnostic Python library for building hybrid keyword, vector, and graph search systems. Applications map their native data to canonical Record values, choose the storage and provider adapters they need, and keep source-specific lifecycle logic outside the kernel.

It supports canonical record search, checkpointed record ingestion, optional local/Postgres/FAISS/provider integrations, and bounded federation across compatible search sources. Breaking changes are documented per release in docs/upgrading.md.

Install

The core package supports Python 3.13 and newer:

pip install andnp-searchkernel

Install optional integrations only when you need them:

pip install "andnp-searchkernel[faiss]"
pip install "andnp-searchkernel[pgvector]"
pip install "andnp-searchkernel[pgvector-psycopg3]"
pip install "andnp-searchkernel[huggingface]"
pip install "andnp-searchkernel[ollama]"
pip install "andnp-searchkernel[markdown]"

First local search

The local composition helper creates durable SQLite-backed stores. This small example uses a deterministic provider so it can run without downloading a model; replace it with a real provider for semantic retrieval.

import asyncio
from datetime import UTC, datetime
from pathlib import Path

from searchkernel import Record, build_local_record_kernel


class DemoEmbeddingProvider:
    model_name = "demo"
    dim = 2

    def embed(self, texts: list[str]) -> list[list[float]]:
        return [[1.0, 0.0] for _ in texts]


async def main() -> None:
    timestamp = datetime.now(UTC)
    composition = build_local_record_kernel(
        Path("records.db"),
        embedding_provider=DemoEmbeddingProvider(),
    )
    record = Record(
        workspace_id="demo",
        source_kind="notes",
        source_id="welcome",
        title="Welcome",
        body="Canonical records can be searched locally.",
        created_at=timestamp,
        updated_at=timestamp,
    )
    composition.keyword_store.index([record])

    outcome = await composition.kernel.search("canonical records", limit=5)
    for result in outcome.results:
        print(result.record.title, result.score)


asyncio.run(main())

The result is a RecordSearchOutcome. Each result retains the complete workspace/source identity and search provenance; degraded execution is reported through outcome.failures and outcome.diagnostics.

Documentation

Start with the documentation map to choose the right guide:

• Getting started — install the package and build a first local index.
• Core concepts — records, identity, ingestion, querying, stores, providers, and readiness.
• Federated search — combine local or HTTP search sources with bounded concurrency and explicit partial-result diagnostics.
• Performance and retrieval roadmap — developer-facing validation evidence, constraints, and future work.

Optional integrations

The core import surface does not require optional providers or backends. The available extras are faiss, pgvector, pgvector-psycopg3, huggingface, ollama, and markdown. See the getting-started guide for the selection rule and the federation guide for the HTTP source adapter.

Validation and releases

CI runs Ruff, Pyrefly, import-linter, the safe test suite, supported Python versions, and selected optional-import checks. Pgvector integration tests need Docker or SEARCHKERNEL_PG_DSN; real-embedding tests are outside the default offline gate. The performance roadmap records what those checks do and do not prove.

Merges to main with release-worthy Conventional Commits are released by the repository workflows. The package and runtime __version__ are checked against each other in CI and before semantic release.

License

MIT. See LICENSE.