Knowledge Intelligence

Enterprise RAG & knowledge intelligence platform: a multi-tenant, permission-aware knowledge platform that grew out of the CSV-AI analysis assistant. Organisations connect documents, spreadsheets, databases and web/code sources; members ask questions and get answers grounded in sources they are allowed to see, with clickable citations down to the page, sheet row or database query.

Status: functional, tested end-to-end on PostgreSQL 16 + pgvector. See IMPLEMENTATION_STATUS.md for exactly what is complete, what is implemented-but-untested, and what is planned.

Demo

A 68-second walkthrough of the real app, with answers from Claude Sonnet 5.5: a cited contract answer, the same question from a member who can't see contracts, a plain-language database question, an agent that waits for an admin's approval, and the live Connections map. Halden Freight is fictional demo data. Watch on YouTube.

Screenshots

Cited answers. Every claim links to its page; the Sources panel shows the passage in context.	Light and dark themes from one token system (docs/DESIGN.md).
Text-to-SQL. A plain-language question becomes validated, read-only SQL; the result is charted.	Approvals. Gated agent tools wait for someone else with db.approve; nobody approves their own request.
Agents resume after approval with the requester's permissions, citing documents and live data.	Connections map. Sources, indexes, models and agents, built from real resources and live events.

Highlights

• Multi-tenant by construction — PostgreSQL Row-Level Security forced on every tenant table, transaction-scoped tenant context, non-superuser DB role enforced at startup, cross-tenant leakage tests.
• Fine-grained access control — 55 permissions, 11 predefined roles + custom roles (escalation-safe), workspace-scoped roles, grants to users/teams/departments/roles/workspaces with deny-wins and expiry, restricted documents, table-limited database grants. Enforced inside every retrieval query.
• Serious RAG — 30+ file formats with page/section/sheet/row locators, 9 chunking strategies, pgvector dense + real BM25 sparse retrieval over the authorized corpus, RRF fusion, feature reranking, MMR, query expansion, parent-child, compression, conversational query rewriting, prompt-injection screening, citation verification, honest refusals. Works offline (extractive answers) or with 10 LLM provider kinds.
• Databases & Text-to-SQL — connector framework (PostgreSQL & SQLite tested; MySQL/MariaDB/SQL Server/Oracle/Redshift/MongoDB beta), schema discovery, table/column selection, sensitive columns, row filters, full/incremental/scheduled sync with deletion propagation, and a sqlglot-based SQL validator with approvals.
• CSV-AI heritage — every uploaded CSV/XLSX/JSON/Parquet becomes a dataset with profiling, deterministic analysis, summaries and natural-language/SQL querying (DuckDB sandbox).
• Connections command center — a live, animated map of the organisation's data ecosystem built from real resources and real-time events (SSE), plus list and data-flow views.
• Operations — PostgreSQL-backed job queue with retries/dead-lettering, webhooks (HMAC-signed), scoped API keys, audit log (append-only), analytics, plan limits, platform administration without tenant content access, Docker Compose with TLS, backups.

Quick start

Docker (single server):

cp .env.example .env        # fill in secrets — see DEPLOYMENT.md
docker compose up -d --build

Local development (no Docker; embedded PostgreSQL):

cd backend && python -m venv .venv && .venv/Scripts/activate && pip install -e ".[dev,local-embeddings]"
python scripts/dev.py                 # API :8000
python scripts/seed_demo.py           # optional demo data
cd ../frontend && npm install && npm run dev   # UI :3000

To get generated answers in development without configuring a provider in the UI, start the API with a file holding ANTHROPIC_API_KEY and/or OPENAI_API_KEY (only those keys, plus ANTHROPIC_WORKSPACE_ID, are read and never printed): python scripts/dev.py --env-file ../.env.keys. Organisations without their own provider then answer with Claude Sonnet 5.5 or GPT-5.4 mini. A local Ollama provider is available too (seed_demo.py --ollama), but small models on a CPU are slow and unreliable.

Documentation

Document	Contents
ARCHITECTURE.md	System design, tenancy, authorization model, ingestion, RAG engine, Text-to-SQL, jobs, events
SECURITY.md	Controls, RAG-specific protections, tests, known limitations
DEPLOYMENT.md	Installation, configuration, operations, backups, upgrades, troubleshooting
docs/API.md	Authentication, errors, endpoint map, examples, webhooks
docs/RAG_EVALUATION.md	Evaluation method and results, incl. live Claude and OpenAI runs
IMPLEMENTATION_STATUS.md	Honest feature matrix
docs/TESTING.md	Backend tests, RAG evaluation and the browser E2E suite
docs/DESIGN.md	Design system
docs/ASSESSMENT.md	Phase 1 assessment of CSV-AI v2
CHANGELOG.md	Release notes

Repository

backend/    FastAPI app, Alembic migrations, tests (164), RAG evaluation
frontend/   Next.js 15 app (design system, 17 sections, chat, Connections map)
deploy/     Postgres init, Caddy (TLS), backup/restore
docs/       API guide, evaluation, testing, design system, assessment

License

MIT — see LICENSE.