Grounded, cited answers
Every response is built from retrieved passages and shows the exact sources — no unsourced guesses.
Self-hosted · Private · Multi-tenant
Your organization's knowledge, answered by your own AI.
A RAG knowledge base and chat assistant that runs on your infrastructure. Upload documents, ask in plain language, and get answers grounded in your own sources — with citations, streaming, and voice. Local-first on Ollama, with optional Groq fallback.
Django
PostgreSQL + pgvector
Ollama
LlamaIndex
FlashRank
Groq
Retrieval
5 sources
Answer
Grounded · cited · streamed
1024-dim
Local embeddings
Hybrid
Vector + keyword retrieval
5 roles
Built-in RBAC
100%
Self-hosted
What it does
Everything you need to turn documents into answers
Ingest, retrieve, and answer — with the guardrails a real organization needs: tenant isolation, citations, and provider choice.
Local-first LLM
Answers run on Ollama on your own hardware. Add Groq cloud fallback only when you want extra speed.
Hybrid retrieval + rerank
Vector search and PostgreSQL full-text search fuse with RRF, then FlashRank reranks the top candidates.
Multi-tenant by design
Companies, roles, and permissions isolate every tenant's documents, models, and API keys.
Ingest anything
PDFs, Word docs, plain text, images with OCR, pasted content, web pages, and Wikipedia topics.
Streaming chat + voice
Token-by-token streaming keeps the UI responsive, and "Hear this" reads answers aloud.
Public chat API
Per-company API keys with rate limits let you embed the assistant in any app or chat box.
Provider switching
Ollama, OpenAI, Anthropic, OpenRouter, or Groq — chosen per tenant, switchable at any time.
Under the hood
From a question to a cited answer
A transparent six-stage pipeline — the same one you can watch in the demo.
1
Ingest
Upload files, paste text, or pull a URL or Wikipedia topic.
2
Extract & chunk
Text is extracted and split into overlapping passages.
3
Embed locally
qwen3-embedding:0.6b turns each chunk into a 1024-dim vector.
4
Store
Vectors live in PostgreSQL + pgvector beside a full-text index.
5
Retrieve & rerank
Vector + keyword search, RRF fusion, then FlashRank.
6
Answer
llama3.2:3b answers from the top chunks, with citations.
Try it yourself
One click in, as any role
The demo ships with a ready-made account for every role — including the platform Admin panel. Pick one and you're signed in.
Admin panel
Platform operator
- All companies
- Django admin
- Global settings
Company Admin
Workspace owner
- Full workspace control
- Roles & providers
- API keys
Manager
Team operations
- Ingest & delete
- Manage members
- Approve signups
Contributor
Content builder
- Upload documents
- Add topics
- Ask questions
Viewer
Read-only access
- Ask questions
- Read answers
- No changes
See the screens
Real pages from the running app
Not mockups — captured from the running app.
Tech stack
Built to run on hardware you already own
Every layer is open source and replaceable. No vendor lock-in, no data leaving your network.
Django + Python Backend
Multi-tenant models, role-based access control, and the admin panel in one framework.
PostgreSQL + pgvector Storage
Vectors and full-text search live in the same database as your business data.
Ollama LLM runtime
Local generation and embeddings on your own GPU or CPU — no per-token billing.
LlamaIndex RAG pipeline
Chunking, embedding, and retrieval orchestration with swap-in components.
FlashRank Reranker
A small cross-encoder that reorders candidates so the best passages reach the model.
Groq / OpenAI Cloud fallback
Optional hosted models per tenant when you need speed beyond local hardware.
See it running in one click
No signup, no credit card. Open the live demo and explore every role.