Graphiti sandbox

Graphiti — temporal knowledge graph for AI agents. Local stack: Ollama (GPU) + FalkorDB + MCP.

Requirements

• Docker + NVIDIA Container Toolkit
• GPU with enough VRAM (default qwen2.5:7b ~5 GB on RTX 3080)

Check GPU in Docker:

docker run --rm --gpus all ollama/ollama nvidia-smi

Quick start

cp .env.example .env
docker compose up -d

First start pulls models via ollama-pull (several GB, one-time). Watch progress:

docker compose logs -f ollama-pull
docker compose logs -f graphiti

Re-pull models manually:

chmod +x scripts/pull-models.sh
./scripts/pull-models.sh

Endpoints

Service	URL
MCP HTTP	http://localhost:8010/mcp/
Health	http://localhost:8010/health
FalkorDB UI	http://localhost:3001
Ollama API	http://localhost:11434

CCT ports untouched: app 8000, redis 6379.

Cursor MCP

.cursor/mcp.json:

"graphiti": {
  "url": "http://localhost:8010/mcp/"
}

Restart Cursor after docker compose up.

Ingest project docs

Index docs/**/*.md into Graphiti graph (entities + facts in FalkorDB, not in repo files):

chmod +x scripts/ingest-docs.sh

# preview
./scripts/ingest-docs.sh --dry-run

# first 3 files (smoke test)
./scripts/ingest-docs.sh --limit 3

# full sync (~94 files, slow on local LLM)
./scripts/ingest-docs.sh

Incremental: .ingest-state.json stores file hashes — re-run skips unchanged docs. Force re-ingest: --force.

MCP search: needs graphiti-core ≥0.29.2 in container (RediSearch fix for cct-backend group_id). Rebuild: docker compose build graphiti && docker compose up -d graphiti.

MCP empty results: FalkorDB graph name must match GRAPHITI_GROUP_ID (default cct-backend). Ingest writes to graph group_id; MCP reads FALKORDB_DATABASE. Both must be cct-backend, not default_db.

Search ingested knowledge via Cursor Graphiti MCP: search_memory_facts, search_nodes.

Models

Default for 10 GB VRAM:

Role	Model
LLM	qwen2.5:7b
Embeddings	nomic-embed-text

Change in .env:

OLLAMA_LLM_MODEL=llama3.1:8b
MODEL_NAME=llama3.1:8b

Then docker compose up ollama-pull --force-recreate and restart graphiti.

Cloud OpenAI fallback

In .env set real OPENAI_API_KEY, OPENAI_API_URL=https://api.openai.com/v1, cloud model names. Remove or scale down ollama service if unused.

Troubleshooting

GPU not used — docker compose exec ollama nvidia-smi inside container.

Ingestion JSON errors — local models weak on structured output; try bigger model or lower SEMAPHORE_LIMIT=1.

Slow first query — Ollama loads model into VRAM on first request.

Stop / reset

docker compose down          # keep graph + models
docker compose down -v       # wipe FalkorDB + Ollama cache