Permission-aware AI search across many companies' data

Data must not cross between client companies, and an HR record must not reach a sales employee. Self-serve ruled out asking clients to restructure data.

Solo: architecture, code, evaluation and the Docker Compose setup.

• Architecture: pgvector as main store and Pinecone behind one interface, so features need no store branches.
• Frontend: Chat screen with cited sources and Right/Wrong marks plus comments for reviewing misses.
• Backend: FastAPI service; PostgreSQL and Dropbox connectors tag chunks with company, groups and shared people.
• Data & AI: Hybrid search, in-process reranking, and model-planned search, SQL or both, guided by business terms.
• Security: Model-written SQL runs read-only behind row-level security and a parser check; uploads parse in time-limited child processes.
• Infrastructure & DevOps: Docker Compose; one command builds each company's schema, indexes and logins from config.
• Testing: 486 automated tests, including isolation on both stores; five before/after evaluations decided what stayed.

• Shared index or schema per company. I chose schemas, so the database, not application code, blocks cross-company reads. Cost: more objects.
• Model-written SQL. I chose a synced copy over live client databases: no load on clients, but data only as fresh as the last sync.
• Too many business terms. 300 terms cost about 38 s per question; a picker cut it to about 12 s, with the right terms for 41 of 41 test questions.

On a synthetic evaluation set: reranking put the right record first for 34 of 40 questions (from 28); SQL planning answered 12 of 13 question types (from 5); terms lifted a wording test from 44 to 51 of 53.