Agentic GraphRAG: Test Scope Analysis

Overview

Dynamic GraphRAG agent that follows information scent: a LangChain create_agent orchestrates vector, BM25, graph traversals, and hybrid RRF based on query intent over Neo4j + Postgres/pgvector. Implements safe vs YOLO modes with deterministic guardrails (PII scrubbing, existence checks) and HITL interrupts. Ingestion uses Docling for PDFs and Tree-sitter for AST-aware chunking, with entity linking into a software QA ontology. Evaluation harness is wired for P@k/MAP/MRR with LangSmith traces; evaluation runs are ongoing.

Key Features

Dynamic agent picks vector, BM25, graph traverse, or hybrid RRF per query scent
Knowledge graph ontology for requirements->code->tests with dual storage (Neo4j + pgvector)
Guardrailed agent modes: Safe (HITL/PII checks) vs YOLO for faster iteration
Docling + Tree-sitter ingestion for PDFs/specs and structure-aware chunking
Evaluation CLI with LangSmith traces and P@k/Recall/MAP/MRR across strategies

Challenges

Entity linking noisy test cases and change requests into the graph ontology
Balancing semantic, lexical, and structural retrieval without adding noise
Maintaining guardrails (PII, HITL) while keeping agent loops responsive

Solution

Modular tool factories (vector_search, graph_traverse, hybrid_search), dual-store writers with retries, and Safe/YOLO toggles backed by PII scrubbing + HITL interrupts. CLI supports chat, ingest, ontology load, and eval to reproduce Ericsson experiments with controllable thinking budgets.

Outcome

Used as the Ericsson master’s thesis artifact; evaluation runs are ongoing with side-by-side baselines (BM25/vector/graph/hybrid vs agentic) instrumented in LangSmith for traceable practitioner feedback.

Project Information

Technologies

Python

LangChain

LangGraph

Neo4j

PostgreSQL

pgvector

BM25

Docling

Poetry

View Live Project View Source Code