Knowhere

Why it's ranked here

The case rests on the whole stack being open under Apache 2.0, not just a client. The server, the background worker and a self-hosting Docker Compose bundle are all public, and the agent endpoint ships in the same process as the HTTP API. The engineering notes show unusual care about correctness: an index is published in the same database transaction as the document, and concurrent writes must converge without losing a committed document. The headline claim, 36 percent better first-try accuracy, comes from the vendor's own internal evaluation with no published method in the fetched text.

What's good

Retrieval scoping is precise and documented: an empty include list returns nothing rather than falling back to everything, and IDs from another user or namespace grant no access. A partially built search index is never exposed; failure rolls the whole publication back and retries. Worker timing is checked at startup, so a misconfigured task limit that would cause stale jobs refuses to boot. Telemetry rules spell out what may leave the machine, and filenames, prompts, emails, IPs and document content are explicitly excluded.

Tradeoffs

Self-hosting is a real deployment, not a library import: it needs Postgres, Redis, a Celery worker queue and a vision-capable model provider. Anonymous usage telemetry to the vendor is on by default and must be switched off with an environment variable. Rate limit rules load only at startup, so changing them means restarting the service. The workspace manifest declares version 0.1.0. The benchmark figures are self-reported internal numbers. The README leads with the paid cloud service and its free credits.

How to use it well

Fit it where an agent has to answer questions over a large set of complex PDFs or slide decks and cite the page it used. Run it with the self-hosted Compose bundle, set the telemetry flag off if you need nothing leaving the network, and point an MCP client at the agent endpoint. Use the include and exclude document lists to fence each query. It is not a lightweight converter for one-off file conversion, and it does not replace a general vector database.

Observed

License

Apache 2.0 (LICENSE)

Language

Python, requires 3.11 or newer (pyproject.toml)

Packaging

uv workspace with API, worker and shared packages; container images on GHCR

Interfaces

HTTP API (FastAPI) plus MCP streamable HTTP endpoint at /mcp (apps/api/main.py)

Infrastructure

Postgres, Redis and Celery workers required (CONTEXT.md, packages/shared-python/shared/core/config/app.py)

Telemetry

Anonymous PostHog telemetry on by default, opt out with TELEMETRY_ENABLED=false

Declared version

0.1.0 in pyproject.toml

Model dependency

External LLM / vision model provider; README suggests deepseek-flash, model-agnostic