Jonathan Woodley

QueryASR

2023

Corporate sustainability reports are hundreds of pages long and nobody reads them end to end. This turns a stack of them into a knowledge base you can interrogate.

The document problem

An annual sustainability report runs to hundreds of pages of prose, tables and footnotes. The information people need from it (a target, a methodology, a restated figure) is real and specific, and full-text search is close to useless for finding it, because the words in the question are rarely the words on the page.

Retrieval rather than summarisation

QueryASR builds a vector index over the documents with LlamaIndex, so a question retrieves the passages that are semantically close to it and the answer is generated from those passages rather than from the model's own recall.

That choice matters for this content specifically: in sustainability reporting a plausible-sounding invented figure is worse than no answer, and grounding every response in retrieved text means the source passage is always there to check.

Where it led

The domain knowledge behind this, what these reports contain and how their figures get constructed, is the same ground the Climate Data Hub covers from the production side.

Role
Sole developer
Stack
LlamaIndex, vector embeddings, Python
Pattern
Retrieval-augmented generation