Agentic GraphRAG

[Naseem77]

Phase 3

Retrieval #274

• 3.2 MCP Server — 8 tools, stdio/sse transport, new mcp/ module
• 3.3 Dynamic Graph Walk — PageRank-weighted traversal, beam search, bidirectional, path scoring
• 3.4 Skill / Tool Library — entity comparison, impact analysis, contradiction detection, gap analysis, timeline reconstruction- [ ] 3.1 Agentic Retrieval (ReAct-Style) — tool loop: search, traverse, Cypher, budget-based stop
• 3.5 Agentic Path Router — one LLM call runs only the retrieval paths a question needs (1–3 of 5), falls back to all 5 on empty/error
• 3.6 Fully Tunable Retrieval — drop hardcoded cutoffs (top_k=15, etc.); agent sets top_k and per-section caps per call, safe defaults for normal flow

Ingestion (next) #275

• 3.7 Configurable Ingestion — agent/caller picks chunker, extractor (LLM vs GLiNER), and resolver per document
• 3.8 Tunable Chunking — expose chunk size & overlap as runtime params
• 3.9 Agentic Extraction — agent picks extraction strategy based on document type/content
• 3.10 Safe Fallback Defaults — standard ingest flow unchanged when nothing is specified

**Testing full pipeline **

• Testing benchmark for novel.
• Testing another pdf document