Graphiti: A Temporal Graph for What Was True When

The Stack

Tested Against

Prerequisites

• Docker
• Python 3.10+
• An OpenAI (or compatible) key for the fenced ingest step

Steps

1. 1

   Start FalkorDB, install graphiti-core, connect, and build indices

   FalkorDB runs in one Docker command. CI then installs graphiti-core[falkordb], waits for the port, connects with FalkorDriver, and builds indices and constraints. A deliberately-dummy OPENAI_API_KEY proves no model is involved in this spine. The container is torn down either way.

   docker rm -f wf-falkordb >/dev/null 2>&1 || true
   docker run -d --name wf-falkordb -p 16379:6379 falkordb/falkordb:latest
   python3 -m venv .venv
   .venv/bin/pip install -q "graphiti-core[falkordb]"
   OPENAI_API_KEY="dummy-key-proves-no-model-call" .venv/bin/python - <<'EOF'
   import asyncio, socket, time
   
   # wait for FalkorDB to accept connections
   for _ in range(30):
       try:
           socket.create_connection(("localhost", 16379), timeout=1).close()
           break
       except OSError:
           time.sleep(1)
   else:
       raise SystemExit("FalkorDB never came up on :16379")
   
   from graphiti_core import Graphiti
   from graphiti_core.driver.falkordb_driver import FalkorDriver
   
   async def main():
       g = Graphiti(graph_driver=FalkorDriver(host="localhost", port=16379))
       await g.build_indices_and_constraints()
       print("graphiti OK: connected to FalkorDB and built indices (no model call)")
   
   asyncio.run(main())
   EOF
   rc=$?
   docker rm -f wf-falkordb >/dev/null 2>&1 || true
   exit $rc

2. 2

   Ingest episodes (the model step, not checked by CI)

   Add episodes with a real key and Graphiti extracts entities and relations with an LLM (OpenAI by default, best with structured-output models). Non-deterministic, so CI never claims it.