Cognik

Self-hosted Agentic RAG knowledge base — an AI agent that captures, retrieves, and writes back knowledge automatically.
No LangChain. No cloud. Your data stays on your server.

简体中文

---

Why Cognik

Most RAG knowledge bases are read-only — you upload docs, the AI answers. When the docs are outdated, answers are wrong.

Cognik is different: the agent writes back. It searches the web, fetches pages, extracts knowledge, and publishes it straight into the RAG pipeline — automatically. Next round, that knowledge is retrievable. No manual uploads, no stale docs.

Self-iterating knowledge loop:

ask → retrieve (BM25 + pgvector + RRF + rerank + CRAG)
    → if gap: web_search → fetch → write back → auto-publish into RAG
    → next question retrieves the new knowledge

Everything runs on your server. PostgreSQL + pgvector + MinIO. No external API calls unless you enable them.

Highlights

• 🔄 Self-iterating Agent loop — search → fetch → write → auto-publish into RAG; semantic de-duplication prevents garbage
• 🔍 From-scratch RAG engine — BM25 + pgvector hybrid → RRF fusion → cross-encoder rerank → CRAG sufficiency assessment (no LangChain)
• 🧠 Six-stage context compression — Tool Result Budget → Snip → Microcompact → HeadAndTail → De-dup → Autocompact (circuit-breaker protected)
• 💾 Memory system — session + global memory, per-turn extraction, cross-session AutoDream review
• 🎫 Ticket workflow — uploads and metadata gaps auto-create review tickets for human-in-the-loop
• 🔐 Private deployment — JWT + RBAC, Docker Compose, supports local llama.cpp inference

Quick start

git clone https://github.com/int2t05/cognik.git
cd cognik

# Start PostgreSQL + pgvector
docker compose -f deploy/docker-compose.yml up -d postgres

# Start backend (:8080)
cd server && go run ./cmd/main.go

# Start frontend (:3000)
cd web && npm install && npm run dev

Open http://localhost:3000 — default account: admin / Admin@123

Tech stack

Layer	Technology
Backend	Go + Gin + GORM
Database	PostgreSQL + pgvector (halfvec + HNSW)
RAG	Self-built Go engine — BM25 (gse) / vector (pgvector) / RRF / cross-encoder rerank / CRAG
LLM	Self-built agent/llm.ChatModel (net/http, any OpenAI-compatible API)
Frontend	Next.js + React + TypeScript + shadcn/ui + Tailwind v4
Deploy	Docker Compose / All-in-One image

Project structure

server/
├── cmd/main.go              # entry
├── internal/
│   ├── agent/              # ReAct loop + tools + memory + compression
│   ├── domain/             # chat / knowledge / ticket / user / system
│   ├── rag/                # self-built RAG engine (14 files)
│   ├── parser/             # document parsing (MinerU + local fallback)
│   └── infra/              # adapter / config / database / storage / middleware
web/src/
├── app/                    # Next.js App Router
├── components/             # shadcn/ui
└── lib/api/                # API clients
docs/                       # PRD / TECH / ROADMAP / API / FLOW
deploy/                     # Docker Compose + All-in-One

Documentation

Doc	Purpose
ROADMAP.md	Product & tech roadmap (V1.0 → V2.0)
PRD.md	Product requirements
TECH.md	Technical architecture & ADR
API/	API contracts (5 modules)
FLOW/	Business flow diagrams (mermaid)

License

MIT