rescore-pipeline

name	rescore-pipeline
description	Rescore pipeline for re-scoring existing JDs from disk. Use when implementing or modifying the rescore CLI subcommand, Rescorer class, or JD file parsing back into JobListing objects.

Rescore Pipeline — Re-Score Existing JDs From Disk

When This Skill Applies

When implementing the rescore CLI subcommand and Rescorer class, or when modifying how JD files are parsed back into JobListing objects.

---

Purpose

Re-score existing JD files against current ChromaDB collections without running a browser search. Use after re-indexing archetypes, adding negative signals, or tuning weights. Completes in seconds instead of 10+ minutes.

---

JD File Format (Input)

Files live in output/jds/ as {rank:03d}_{company_slug}_{title_slug}.md. For full format specification, parsing rules, and edge cases, see jd-file-format.md.

---

Data Flow

output/jds/*.md
    → parse metadata + body → list[JobListing]
    → _ensure_indexed() (auto-index if collections empty)
    → for each listing:
        → scorer.score(full_text)           → ScoreResult (fit, archetype, history, negative)
        → comp_parser.parse(full_text)      → comp fields
        → compute_comp_score(comp_max)      → comp_score
    → ranker.rank(scored, embeddings)       → ranked list + summary
    → export: results.md, results.csv, jd files (with new scores/ranks)

Decision Skipping

• Same as search: check DecisionRecorder.get_decision(external_id) before scoring
• Skip decided listings unless --force-rescore is passed
• Skipped count appears in run summary

---

CLI Interface

# Re-score all undecided JDs
python -m jobsearch_rag rescore

# Re-score everything including decided
python -m jobsearch_rag rescore --force-rescore

Exit Conditions

Condition	Behavior
output/jds/ missing	Print actionable message + exit
output/jds/ empty	Print actionable message + exit
No valid JD files parsed	Print warning + exit
Collections empty	Auto-index before scoring (same as search)

---

Module Location

pipeline/rescorer.py — standalone module, not bolted onto runner.py.

Shares with runner:

• Scorer, Ranker, Embedder, VectorStore, DecisionRecorder
• parse_compensation(), compute_comp_score()
• All exporters (MarkdownExporter, CSVExporter, JDFileExporter)

Does NOT share:

• Browser session management
• Adapter layer
• Rate limiting / throttling

---

handle_rescore in cli.py

Pattern mirrors handle_search:

1. load_settings()
2. Configure file logging
3. Create Rescorer(settings)
4. asyncio.run(rescorer.run(force_rescore=args.force_rescore))
5. Print summary (same format as search)
6. Run all three exporters