Graph analytics for people who forgot their algorithms textbook

What it does

YiGraph ingests raw text, logs, or tables, extracts entities and relationships, and builds a structured graph. You type a business question in natural language; the system plans the analysis, invokes actual graph algorithms, and returns a traceable report. It is pitched at fraud detection, supply-chain risk, and money-laundering investigations—any domain where the connections matter more than the nodes.

The interesting bit

The project uses an Analytics-Augmented Generation (AAG) framework, which is a polite way of saying the LLM is kept away from the calculator. The model handles intent and narration, but the heavy lifting—centrality scores, community detection, pathfinding—is delegated to a vetted library of over 200 algorithms. This keeps the output reproducible and stops the agent from confidently inventing a shortest path.

Key highlights

• Ships with 200+ graph algorithms (PageRank, Louvain, k-core, etc.) backed by NetworkX and Neo4j.
• Three operating modes: fully autonomous “Normal,” collaborative “Interactive,” and bring-your-own-plan “Expert.”
• Task-aware graph construction: it builds only the subgraph relevant to the question rather than dumping everything into one giant hairball.
• Accepts both structured graph data and unstructured text, automatically extracting entities and relationships from documents and logs.

Caveats

• The documentation currently targets Neo4j 3.5.25 and Java 8/11, which are several major versions behind the current release line.
• The README claims both 11 and 21 algorithm categories in different sections, so the exact taxonomy is unclear.

Verdict

Worth a look if you need graph analytics but would rather describe the problem in English than write Cypher queries. Skip it if you already have a tuned graph pipeline and just want a thinner LLM wrapper.