mnemiq

Why it's ranked here

Apache 2.0, a Python 3.12 package with three front doors: a command line, an HTTP API behind a browser workbench, and an MCP (Model Context Protocol) server exposing three tools to agents. The design choice that matters is where trust sits. The model proposes, deterministic checks decide, and access grants nothing until a role is named. The evaluation harness reports right, declined and wrong as separate numbers, and grading lives in a separate repository so the engine does not mark its own homework. The caveat is maturity: the package declares a pre-release development version, so expect interfaces to move.

What's good

Code meanings follow a stated precedence: a standard vocabulary fills gaps, labels found in the data beat it, and an operator dictionary beats everything. A vocabulary only attaches to a column when values, name and eligibility all agree, and two competing matches bind neither. Private definitions are hidden by default; only vocabularies explicitly marked public are shown to everyone. A local-only switch refuses to start if any model endpoint is not a loopback or private-range address, and it does no DNS lookups it cannot trust. Writes exist but are refused unless both a write grant and a deployment setting allow them, and schema changes never run.

Tradeoffs

First install pulls about 230 MB of dependencies, mostly DuckDB, PyArrow and the OpenAI client. Enrichment needs a live model and prints nothing until each table finishes, which reads as a hang on large schemas. The local-only check skips database connection strings by design, so those need verifying by hand. Air-gapped runs are not turnkey: DuckDB downloads its search and source extensions on first use unless the cache is pre-seeded. The browser workbench requires a separate Node build step. The highest-precision answer mode runs five candidates plus a judge, roughly six times the cost of the default.

How to use it well

Built for a data team deciding whether natural-language querying will hold up on its own warehouse before buying a hosted product. Follow the project's own advice: take one schema slice, write 20 to 30 real questions tagged by whether column names, a definition, or nothing should answer them, then compare runs with enrichment on and off and two verifier thresholds. Always pass a role, since an empty one returns deferrals that look like bugs. Supply a dictionary for internal codes. It is not a migration tool, and Postgres, SQLite, Oracle, Snowflake and Databricks are the sources the repository names.

Observed

License

Apache License 2.0, copyright Agentic Fabriq

Language

Python, requires Python 3.12 or later

Packaging

Built with hatchling; installs a mnemiq console command; uv is the documented install path

Core dependencies

DuckDB, psycopg, the OpenAI client, sqlglot, PyArrow, pydantic and the MCP SDK

Optional extras

Oracle, Snowflake, Databricks, an HTTP server (FastAPI, uvicorn) and ontology digestion (rdflib)

Interfaces

CLI, HTTP API with an SSE chat endpoint, and an MCP server over stdio exposing db_read, db_write and get_schema

HTTP API

Serves /v1/ask, /v1/chat and /v1/schema and returns deferrals as HTTP 200 rather than errors

Model endpoints

Any OpenAI-compatible /v1 endpoint, including local servers such as vLLM, Ollama and llama.cpp

Ontology input

Accepts SKOS concept schemes and OWL class hierarchies from TTL files

Testing

pytest, with markers for integration (live database) and live_llm tests

Local services

Docker Compose file defines a Postgres 16 service

Contract

Snapshot and certified-record contract published as a JSON Schema file in the repository