LLM reference segmenter and citation engine - #702

Draft

de-code wants to merge 12 commits into

mainfrom

llm-reference-segmenter-engine

Commits on Aug 20, 2026

Add an opt-in LLM engine for the reference models
de-code
committed
d4dce70
Add optional tracing to the LLM engine
de-code
committed
5896bc4
Batch citation references into one request
de-code
committed
11c6476
Add an evidence response shape for the reference segmenter
de-code
committed
a8c88a9
Stop the citation model returning the batch marker as a field
de-code
committed
6ded2c5
Drop an overlapping field rather than failing the document
de-code
committed
c709e40
Drop a claim the engine cannot honour rather than failing the document
de-code
committed
ba9ec38
Run citation batches concurrently, and expose both knobs in the config
de-code
committed
d67ceba
Serve qwen3.5-9b from Venice rather than SiliconFlow
de-code
committed
227113b
Make the LLM engine usable from the benchmark workflow
de-code
committed
e9fa377
Move the llm guard test out of tests/
de-code
committed
350fc65
Narrow the profile lookup explicitly for mypy
de-code
committed
cc804bd