Indexing The Graph + RAG: AI Copilot for On-Chain Activity

Frequently Asked Questions

01.Why pair The Graph with RAG instead of querying an RPC node directly?

An RPC node returns raw logs and bounded ranges, and eth_getLogs caps responses at 10,000 results or roughly a 2,000-block window on major providers. A language model cannot scan an entire chain history that way. The Graph indexes raw activity into a GraphQL schema, and RAG retrieves the relevant slice so the model answers from grounded data rather than guessing.

02.How does The Graph differ from a vector database in this stack?

The Graph is the indexing and structured-query layer that turns raw chain events into deterministic, schema-shaped records. A vector database such as pgvector stores embeddings of those records and powers semantic retrieval. They are complementary: The Graph supplies ground-truth state, and the vector store supplies the nearest-neighbour lookup that feeds the model context window.

03.What are the main failure modes of an on-chain AI copilot?

The common failures are stale or reorged data feeding the retriever, embedding drift when token metadata changes, prompt injection through attacker-controlled on-chain text fields, and the model fabricating addresses or balances when retrieval returns nothing. Every answer should cite the subgraph record it came from, and write actions need on-chain circuit breakers and human approval.