Features

A complete AI knowledge system with retrieval, verification, intelligent agents, and a knowledge graph — all running locally.

Retrieval
Hybrid Search
BM25s with stemming + ChromaDB vector search. Semantic chunking preserves context across document boundaries.
Verification
Streaming Verification
Real-time claim verification as the LLM responds. Four claim types: evasion, citation, recency, and ignorance detection.
Quality
Self-RAG Validation
Self-reflective retrieval-augmented generation. The system validates its own retrieval quality before responding.
Routing
Smart Model Routing
Three-way routing with capability-based model scoring. Direct-to-OpenRouter chat proxy. Proactive model switching on ignorance detection.
Ingestion
Universal Ingestion
PDF, Office, email, ebooks, structured data. Watch directories for automatic ingestion. Drag-and-drop in the UI.
Graph
Knowledge Graph
Neo4j-backed relationships between artifacts. Domain taxonomy, quality badges, tag vocabulary with typeahead.
Chat
Context-Aware Chat
Corrections, token-budget KB injection, semantic dedup. Smart KB suggestions surface relevant context automatically.
Integration
18 MCP Tools
Full Model Context Protocol integration. Query, ingest, search, manage taxonomy, and more — all via MCP.
Reliability
Circuit Breakers
All Bifrost and Neo4j calls protected by circuit breakers. Graceful degradation when services are overloaded.
Agent System

9 intelligent agents

Specialized agents collaborate to ensure high-quality retrieval, verification, and knowledge management.

Query

Orchestrates RAG retrieval with context-aware KB injection

Triage

Routes queries to the most capable model

Curator

Manages knowledge quality and metadata enrichment

Rectify

Detects and corrects outdated or conflicting knowledge

Audit

Monitors system health and data integrity

Maintenance

Handles cleanup, deduplication, and optimization

Hallucination

Validates claims against your knowledge base

Memory

Manages long-term conversational memory

Self-RAG

Self-reflective retrieval for higher accuracy

Architecture

Six core services working together on a shared Docker network.

MCP Server

Port 8888 · FastAPI / Python 3.11

Bifrost Gateway

Port 8080 · Semantic intent routing

ChromaDB

Port 8001 · Vector database

Neo4j

Port 7474 · Graph database

Redis

Port 6379 · Cache + audit log

React GUI

Port 3000 · React 19 + Vite + nginx