If you’ve wired an internal agent to a warehouse, you know the loop: toy prompt works, then a slightly vague question invents a JOIN, remaps “tickets” onto the wrong table, or picks qty vs unit_amount. Green CI. Garbage query. you can try with just logging in. free Sandbox on the compile API. also we split the digits for params — “last 30 days” does not become $30. binds have to match the prompt. no guessing qty vs unit_amount. also half of this is in a pip lib for getting wet. lol that half is PrismSchema (Apache-2.0). DDL dump / .sql / local read-only Postgres or MySQL → JSON/YAML contract. it never talks to the compile API. never sends your DSN off-box. pip install "prismschema==0.1.0" https://github.com/insightitsGit... other half is PrismSQL (hosted). upload the contract. model never writes SQL or ON. closed Query IR → compiler (windows, compares, ORDER BY on declared columns, JWT email as a bind, dialect printer). you get sql_template + params (%s). hosted rows always []. you run it. we never take a warehouse DSN. if the prompt is honest-but-incomplete it stops. HTTP 422, status: clarify, candidates[] — e.g. line items over 5 when both qty and unit_amount exist. unknown tables refuse. not remapped. Sandbox free (PostgreSQL). more dialects on paid plans. no published benches. compiler 1.6.5. https://www.insightits.com/produ... https://www.insightits.com/docs/... how are you stopping the model from writing the JOIN — allowlists, compile, or hoping the prompt holds?

English TO SQL

2 layer retrievals no just cosine similarity

2 layer retrievals no just cosine similarity