g1-brain 🧬

The memory and intelligence layer — graph memory, 4-stream knowledge retrieval, and structured reasoning for the Generate One platform.

---

✨ Overview

g1-brain provides the platform's memory and intelligence layer. It combines graph-based persistent memory (Graphiti + Neo4j) for context across sessions, vector search (Qdrant) with 4-stream Reciprocal Rank Fusion for document retrieval, and structured reasoning tools for multi-step analysis. This is what gives the Generate One platform long-term memory, deep knowledge retrieval, and the ability to reason over facts over time.

Two compose stacks are managed from this repo: memory-stack (Neo4j, Graphiti API, Redis, FalkorDB) and knowledge-mcp (knowledge-mcp server, async worker, Valkey queue).

---

🏗️ Architecture

graph TD
    subgraph "Memory Stack (z0ww84wkwwss8sw4kgsw88gc)"
        Neo4j["Neo4j 5.26\n(graph DB, port 7474/7687)"]
        GA["Graphiti API\n:8000 → memory.generate.one"]
        Redis["Redis 7-alpine\n(Graphiti cache, 256MB LRU)"]
        FalkorDB["FalkorDB\n(graph query engine)"]
        GA --> Neo4j
        GA --> Redis
    end

    subgraph "Knowledge Stack (kp3basi7wsdztrq1fgm7t543)"
        KM["knowledge-mcp\n:8000"]
        KW["knowledge-worker\n(async ingest pipeline)"]
        KV["Valkey 9.0.1\n(task queue)"]
        KW --> KV
        KM --> KV
    end

    subgraph "MCP Tools"
        GMCP["graphiti-mcp"]
        RMCP["reasoning-tools"]
        LMCP["logic-lm"]
        CMCP["cms-tools\n(Directus)"]
    end

    GMCP --> GA
    KM -.-> Qdrant["g1-llm / Qdrant"]
    KW -.-> Qdrant
    KW -.-> LiteLLM["g1-llm / LiteLLM"]

    Client["MCP Clients"] --> GMCP
    Client --> KM
    Client --> RMCP
    Client --> LMCP
    Client --> CMCP

---

📦 Services

Memory Stack (compose/memory-stack.yml)

Service	Image	Port	Description
graphiti-api	Custom build	8000	REST API for Graphiti graph operations → memory.generate.one
neo4j	neo4j:5.26-community	7474, 7687	Graph database for entity/relationship storage
redis	redis:7-alpine	6379	Graphiti internal cache (256 MB LRU)
falkordb	falkordb/falkordb:latest	—	Graph query engine (Cypher-compatible)

Knowledge Stack (docker-compose.yml — service kp3basi7wsdztrq1fgm7t543)

Service	Image	Port	Description
knowledge-mcp	git.generate.one/generate-one/knowledge-mcp:latest	8000	MCP server with 7 search/ingest tools
worker	git.generate.one/generate-one/knowledge-worker:latest	—	Document ingest pipeline (OCR, chunking, embedding)
knowledge-valkey	valkey/valkey:9.0.1-alpine	6379	Task queue for async ingestion

---

🔍 Knowledge Search: 4-Stream RRF

Queries are processed through 4 parallel retrieval streams, fused via Reciprocal Rank Fusion:

1. Dense vectors — Semantic similarity via Qdrant
2. BM25 sparse vectors — Keyword matching with IDF weighting
3. Graphiti — Entity/relationship graph traversal
4. Reranker — Cross-encoder rescoring (mxbai-rerank-large-v2)

Post-fusion temporal decay (30-day half-life) prioritizes recent information.

---

🔧 Configuration

Variable	Description	Default
NEO4J_AUTH	Neo4j authentication (neo4j/password)	—
GRAPHITI_API_KEY	API key for Graphiti REST API	—
LITELLM_API_KEY	LiteLLM key for LLM operations	—
VLM_OCR_ENABLED	Enable VLM-based OCR for scanned PDFs	true
VLM_OCR_MODEL	Model tier for OCR	g1-vlm
CLASSIFY_LLM_MODEL	Model for document classification	g1-llm-turbo
CONTEXTUAL_CHUNKING_ENABLED	LLM-enriched chunk context	true
CONTEXTUAL_LLM_MODEL	Model for chunk enrichment	g1-llm-micro
GRAPH_DB_PASSWORD	Neo4j database password	—
FALKORDB_PASSWORD	FalkorDB auth password	—

---

🏠 Multi-Tenant Isolation

Service	Parameter	Default
graphiti-mcp	group_id	main
reasoning-tools	group_id	main
knowledge-mcp	tenant_id	shared

Group ID format uses dashes only (colons rejected): main, tenant-{id}, org-{org_id}-user-{user_id}

---

🚀 Quick Start

# Memory stack (managed by Coolify)
cd /data/coolify/services/z0ww84wkwwss8sw4kgsw88gc
docker compose up -d

# Knowledge stack (locally-built images)
cd /data/coolify/services/kp3basi7wsdztrq1fgm7t543
docker build -t git.generate.one/generate-one/knowledge-worker:latest worker/
docker build -t git.generate.one/generate-one/knowledge-mcp:latest .
docker compose up -d

# Health checks
curl https://memory.generate.one/health

Note: knowledge-mcp and knowledge-worker use locally-built images with pull_policy: never. Always rebuild before docker compose up -d.

---

🔗 Dependencies

Depends on:

• g1-llm — Qdrant (vector storage), LiteLLM (inference for classification, enrichment, query rewriting)

Depended on by:

• g1-mcp — graphiti-mcp and reasoning-tools connect to memory/knowledge backends
• g1-gpt — LibreChat file search routes through knowledge-mcp RAG compatibility layer

---

🔗 Related Repos

Repo	Relationship
g1-llm	Qdrant + LiteLLM for embeddings and inference
g1-mcp	graphiti-mcp, reasoning-tools hosted in mcp-stack
g1-gpt	LibreChat file search via knowledge-mcp
g1-core	Valkey for task queueing

---

🛡️ Part of Generate One

Generate One — AI infrastructure that answers to you.

Self-hosted, sovereign AI platform. generate.one

Licensed under AGPL-3.0.