Applied AI / LLM
Agentic Graph RAG
PythonFastAPIFalkorDBLangChainStreamlitSSE
An end-to-end agentic GraphRAG stack: ingest almost any file into a FalkorDB knowledge graph, then ask questions in natural language. A tool-calling agent inspects the live graph schema, plans and executes read-only Cypher, self-corrects on errors, and streams its full reasoning trace — grounding every answer in the graph.
Architecture
Highlights
- Model-agnostic: every LLM call goes through one provider layer, switchable between OpenAI, Anthropic, Ollama, or any OpenAI-compatible endpoint with two environment variables — and swappable at runtime from the UI.
- Modern agent loop: native tool calling with conversation-state accumulation, a structured-JSON fallback for local models, read-only Cypher validation, error self-correction, and capped retries.
- Memory that scales: recent turns replay verbatim; older turns fold into a rolling LLM-generated summary so long chats keep context without unbounded prompts.
- Ingest any shape: tabular files become a typed entity graph via inferred schema; documents are chunked with LLM-extracted entities and relationships.
- Streaming API: the FastAPI backend streams agent events (thoughts, tool calls, Cypher, answers, graph payloads) over SSE to a Streamlit UI with live traces and an interactive graph view.