SRag
RAG with Cognitive Memory
A knowledge system that extracts entities, facts, and relationships from every document. Detects contradictions, self-heals outdated knowledge, and provides layered search combining knowledge context with graph-boosted retrieval.
Everything you need for production RAG
A complete framework with cognitive memory that handles ingestion, knowledge extraction, and intelligent retrieval.
Cognitive Layer
Knowledge Graph
Entities and relationships extracted from every document. 5-stage entity resolution with 4-way alias matching, type normalization, and all-LLM smart merge across documents.
Self-Healing Memory
Atomic facts tracked over time. Contradictions, refinements, and retractions detected automatically — outdated facts are superseded and entity descriptions refreshed.
Cognitive Search
Layered response: knowledge context (entities, facts, relationships) plus graph-boosted source chunks. Hybrid entity and fact search fused together for brain-like recall.
Entity Resolution
Two-phase merge: parallel candidate finding via embedding similarity and 4-way alias matching, then sequential merge via union-find graph. Every decision confirmed by LLM with full evidence.
Core Engine
6-Stage Retrieval
Query expansion, vector similarity, BM25 keyword search, Reciprocal Rank Fusion, MMR diversity filtering, and LLM reranking — a full retrieval pipeline in every search.
Dual-Pipeline Ingestion
Two independent pipelines: chunks are ready immediately, while the cognitive pipeline extracts knowledge in the background. Fully resumable with per-chunk persistence.
Smart Chunking
Four strategies — token, recursive, semantic, and no-split. Content is sanitized before chunking, and each document supports per-document configuration with re-ingestion.
Supabase + pgvector
PostgreSQL with halfvec embeddings for 50% storage savings. BM25 via tsvector, HNSW indexes, scope-aware indexing, and semantic deduplication.
Integration
16 MCP Tools
Model Context Protocol server with streamable HTTP transport. One-click Cursor setup, automatic scope injection, and tools for search, ingest, entities, facts, and timeline.
Full Observability
Every ingestion and cognitive step logged with timing and metrics. Live activity feeds via Supabase Realtime, processing timelines, and per-stage debug info in search results.
Project Isolation & Sharing
Every document belongs to a project with an owner. Row Level Security enforces access in the database itself — invite members as viewers or editors, share single documents, and query across everything you can access.
Flexible Providers
OpenAI, Azure, Ollama, and AWS Bedrock embedding providers via LangChain. Configurable LLM models for extraction, reranking, and query expansion.
How it works
Two independent pipelines turn raw content into searchable knowledge.
Ingest
Send content via REST API, MCP tools, or the admin UI. Documents are validated and accepted in milliseconds.
Process
Two pipelines run independently — chunks are ready immediately, knowledge extraction continues in the background.
Query
Layered search returns knowledge context (entities, facts, relationships) alongside graph-boosted source chunks.
See it in action
From raw documents to brain-like memory — walk through how the cognitive engine gives AI agents the ability to recall facts, track changes, and make connections across your entire knowledge base.
Sarah Chen (CTO) led the launch of Project Atlas using DuckDB...
We migrated from ClickHouse to DuckDB for analytical queries...
Sarah Chen was promoted to CEO effective April 2025...
Feed any content. Chunks are searchable immediately while knowledge extraction runs in the background.