Seulanga RAG
Autonomous Offline-First Knowledge System — Hybrid adaptive search, adaptive AI memory, MCP server
Seulanga RAG
v3.2 Production
Seulanga RAG is an offline-first personal knowledge system that processes documents through an 8-stage pipeline — ingestion, chunking, embedding, indexing, hybrid adaptive search, reranking, agent memory, and adaptive loop. Exposes MCP server (stdio + HTTP) for AI agent integration.
Pipeline Architecture
1
Ingestion — file, dir, URL, paste, raw text
→
2
Chunking — paragraph, semantic, fixed, sentence, hybrid
→
3
Embedding — TF-IDF, BM25, LLM, ONNX, hash (with fusion)
→
4
Indexing — vector, inverted, graph, hybrid store
→
5
Search — 5 strategies: MFK, BFS, HR, HB, HS (auto-weighted)
→
6
Reranking — cross-encoder, MMR, diversity
→
7
Memory — 12-slot STM → graph LTM with CRDT
→
8
Agent Loop — self-heal, self-verify, rewrite queries
MCP Server
Model Context Protocol server with 6 tools, available via stdio or HTTP (port 3100).
ingest
Semantic chunking + multi-embedding fusion → store
search
Hybrid adaptive search with reranking
remember
Agent memory CRUD (STM + LTM)
status
System health + stats + memory dashboard
forget
Memory pruning with importance decay
graph
Knowledge graph entity + relationship query
Embedding Methods
TF-IDF — Term frequency-inverse document frequency
BM25 — Probabilistic ranking with saturation
LLM — Provider embeddings via Hermes/Jeumpa
ONNX — Local ONNX model inference
Hash — Fast locality-sensitive hashing
Search Strategies
MFK — Multi-Field Keyword
BFS — Best-First Search (semantic)
HR — Hybrid Ranked
HB — Harmonic Blended
HS — Hybrid Summarization
Self-Healing Agent
The system includes an agent-in-the-loop that monitors its own outputs and health:
- Self-Verify — Checks answer relevance and correctness against sources
- Query Rewrite — Automatically reformulates ambiguous queries
- Memory Healing — Detects and repairs gaps in agent memory
- Pipeline Health — Monitors each stage; auto-fallback on failure
- Cryptographic Signing — Memory entries signed with Ed25519