docs/ai/COMMIT_CONTENTS.md

Initial Commit Contents

This file defines the deployed contents of the lightweight ENCODE Space and its separate private runtime-artifact bucket.

Application and deployment

• README.md — project overview and Hugging Face Space metadata.
• docs/ai/COMMIT_CONTENTS.md — this auditable release inventory.
• Dockerfile — CPU-only backend container listening on port 7860.
• netlify.toml and deploy/ — Netlify build/Drop package that proxies the static frontend's API requests to the Hugging Face Space backend.
• .dockerignore — excludes development-only files from container builds.
• .gitignore — excludes data/ and the generated parts of outputs/.
• src/requirements.txt — pinned Python runtime dependencies.

Backend

• src/backend/app.py — FastAPI application and HTTP endpoints.
• src/backend/codesearch.py — medical-code retrieval.
• src/backend/retriever.py — phenotype retrieval using Vidul's fine-tuned embeddings.
• src/backend/encode.py — annotation and response handling.
• src/backend/graph.py — phenotype detail graph construction.
• src/backend/models.py — registry of the retrieval models the UI may offer.
• src/backend/rxnav.py — conservative, cached RxNAV/RxNorm completion for medication mappings that are absent from packaged outputs.
• src/backend/__init__.py — backend package marker.

Frontend

• src/frontend/index.html — browser application shell.
• src/frontend/app.js — search, review, and annotation behavior.
• src/frontend/styles.css — application styling.

Private bucket: runtime model and indexes

• outputs/models/bge_ft_va/ — Vidul's fine-tuned model weights, tokenizer, and configuration.
• outputs/indexes/bge_ft_va/icd_index/ — diagnosis FAISS index and metadata.
• outputs/indexes/bge_ft_va/med_index/ — medication FAISS index and metadata.
• outputs/indexes/bge_ft_va/ndc_index/ — NDC FAISS index and metadata.
• outputs/indexes/bge_ft_va/labchem_index/ — LOINC/lab FAISS index and metadata.
• outputs/indexes/bge_ft_va/cpt_index/ — procedure FAISS index and metadata.
• outputs/embeddings/bge_ft_va/code.npy
• outputs/embeddings/bge_ft_va/code_meta.json
• outputs/embeddings/bge_ft_va/config.json
• outputs/embeddings/bge_ft_va/metadata.npy
• outputs/embeddings/bge_ft_va/metadata_ids.json

Private bucket: runtime phenotype records

• outputs/canonical/algorithm_components.jsonl
• outputs/canonical/canonical_phenotypes.jsonl
• outputs/canonical/cipher_links.jsonl
• outputs/canonical/code_descriptions.jsonl
• outputs/canonical/lab_loinc_map.jsonl
• outputs/canonical/loinc_terms.jsonl
• outputs/canonical/med_ingredient_map.jsonl
• outputs/canonical/rxcui_ingredient_map.jsonl

Verification

• tests/test_retrieval.py — focused tests for the retained dense-retrieval path.
• tests/test_rxnav_and_deployment.py — RxNAV provenance and Netlify Drop deployment checks.

Explicitly excluded

Everything under data/ (raw VA exports, the original application bundle, and the LOINC and RxNorm distribution archives) and everything the pipeline writes to outputs/ — training, evaluation, reranker, facet, profiling, and cached results, generated demos, and annotations — is excluded from the Space upload, along with local environments, credentials, and editor settings.

The release contains one retrieval path: Vidul's fine-tuned model, prebuilt embeddings, and exact dense search. Large runtime assets are stored in hf://buckets/hiasgnpsadgd/encode-artifacts and mounted read-only at /home/user/app/outputs; they are not committed to the Space repository or copied into the Docker image.