Choosing a method: baseline vs structured¶

When do you pay for graph extraction, and when are raw chunks enough?

This page is a decision aid, not an evaluation framework. It does not define recall metrics, QA pairs, or CI scores. For a same-corpus walkthrough of answers, latency, and token use, run the companion script on examples/en/tesla.md once it is in examples/en/methods/ (chunk_vs_graph_rag.py). That script is a manual demo — CI will not run it.

The catalog of algorithms lives in Methods. How to call them is in Using Methods.

Two defaults¶

Need	Start with	Why
Lookup, quote the original wording, or locate a passage	`chunk_rag`	Ingest is embed-only. No LLM extraction cost. Retrieval returns the chunks themselves.
Multi-hop or relationship-dense questions	`graph_rag`	Extraction builds nodes and edges so answers can follow links, not only lexical overlap.

chunk_rag shipped as the zero-extraction baseline. Use it first; pay for graph_rag (or another structured method) only when that extra ingest cost changes the answers you care about.

When chunks are enough¶

Prefer chunk_rag when:

The question is lookup-style: “What year…?”, “Quote the clause…”, “Where does the document say…?”
You need the original wording, not a synthesized graph fact.
The corpus is large and ingest LLM cost is the constraint.
You are deciding whether extraction is worth it at all — this is the baseline.

Ingest path: chunk → embed → index. Query path: retrieve chunks → (optional) chat over those chunks.

When graph extraction pays off¶

Prefer graph_rag (or another graph/hypergraph method) when:

The question is multi-hop: “Who funded the company that acquired X?”
The question is relationship-dense: rivals, successors, ownership, causation.
You will reuse the graph (export, scoped graph search, visualization), not only ask once.

You pay an LLM at ingest to extract structure. That cost is wasted if every question could have been answered by pointing at a chunk.

Same corpus, same provenance¶

chunk_rag and graph_rag share the provenance machinery: source ledger, he tag, scoped he search / he talk, and per-document rollback. A comparison on one corpus (examples/en/tesla.md) is apples-to-apples — the difference is extraction, not tagging or rollback.

See Source Attribution & Provenance.

This is not an eval harness¶

This page does not define hit-rate metrics, expected spans, or a benchmark set. Hash embeddings in CI would make recall numbers meaningless. A separate RFC would be needed before any harness — that work is out of scope here.