Skip to content

feat: implement hallucination detection #12

@DelusionalSimon

Description

@DelusionalSimon

Add an internal verification step to cross-check the LLM's claims (gene names, locus tags) against the input annotation file.

Tasks

  • Create a new helper function that runs after the LLM call.
  • This function will parse the potential_features and check if the supporting_genes and locus_tags exist in the original data.
  • Define the additions to the JSON scheme (how explicit to be in the tag)
  • The function will add a new key to each feature in the JSON, ("verification_status").
  • If a gene or locus tag is not found or gets mismatched, it will tag it appropriately ("verification_status": "possible hallucination").
  • The helper function for the human-readable .txt report will add some notes on detected hallucinations in the output
  • Test thoroughly with known-good and known-bad data.
  • Update documentation to explain the verification status tags.
  • (Optional) look into more advanced hallucination detection techniques (RAG)

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions