Non-LLM extraction backend for graph indexing: 1000× cheaper, schema-bounded — where would it fit in GraphRAG?

[Jeffrey Romero Del Val (DeliVali)]

Hi — long-time reader of this project's design docs. I built something adjacent to GraphRAG's indexing stage and I'd genuinely like this community's take on where (or whether) it fits.

The problem I was attacking

For high-volume corpora, the entity/relationship extraction stage is where GraphRAG's cost concentrates: every text unit goes through LLM calls, and re-indexing multiplies it. For always-on ingestion (agent memory, log streams) that cost structure makes "index everything" impractical.

What I built

A small non-autoregressive decoder (~37M params) that maps a sentence embedding (SONAR, Meta) directly to graph triples — entities + typed relations in one forward pass. No generation, no JSON parsing, no retries: malformed output is structurally impossible.

• ~0.013 ms/sentence on a consumer RTX 5070 (trained end-to-end on that same 12GB card)
• Held-out triple-F1: 1.000 on tool-call structures, 0.827 on prose with entity candidates, 0.651 open-vocab
• Cross-lingual for free (the embedding space is language-agnostic): Spanish input against an English entity vocabulary resolves to the same nodes

Repo (Apache-2.0, weights on HF, reproducible): https://github.com/DeliVali/cogito-estella

The honest tradeoff vs LLM extraction

This is not a drop-in replacement for GraphRAG's extractor. LLM extraction gives you open entity types, descriptions, and claims; mine works over a bounded vocabulary (20k entities / 60 relations currently) and emits bare typed triples. Where I think it fits:

1. Tiered indexing: cheap pass over everything, LLM pass only where the cheap pass finds density worth enriching
2. Re-indexing / incremental updates: the cost case where per-token pricing hurts most
3. Domains where a curated schema is a feature (compliance, observability), not a limitation

Questions for this community

1. Has anyone experimented with non-LLM or hybrid extraction backends for the indexing stage? What broke?
2. For those running GraphRAG at scale: what does extraction actually cost you per, say, 1M text units? I want to check my "1000×" math against real deployments.
3. If a tiered/hybrid backend were pluggable, what would the interface need to produce to be useful downstream (community detection, summarization)?

Not trying to sell anything — the thing is free and I'm one person. I'm trying to figure out if this direction is useful to people other than me, and this is the community that would know.