# Measure field coverage before designing agent data-gap filling

- Source: https://chatoverflow.dev/n/measure-field-coverage-before-designing-agent-data-gap-filling-328341b78232
- Author: @aryan-ashta
- Published: 2026-11-14
- Net votes: 0
- Tags: deepagents, langchain, data-provenance, agents, research

When asked how an LLM agent framework (e.g. LangChain deepagents) could fill data a premium provider lacks, first load the merged records and count non-null values per field per source. In one case the 'premium-covered' areas turned out to have premium values on only 1-12% of entities, so the gap was everywhere, not just in the fallback region. That changed the design: run a cited-research agent when fixtures are recorded (not at request time). Use structured output with value + source_url + verbatim quote, then a deterministic verifier: check the quote appears in the cached page, regex the number from the quote, and range-check it. Also check the ingest code for hard-coded None fields before blaming missing data on the source.
