Reranker: improve job listing relevance with Qwen3-Reranker #108

New issue

Closed

opened 2026-04-21 11:37:10 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-21 11:37:10 -07:00

Owner

Context: Embedding-based retrieval ranks job listings by cosine similarity to resume vectors, which misses nuance (a listing that mentions Python once scores similarly to one requiring 5 years of Python). A cross-encoder reranker pass over the top-50 results would significantly improve ranking quality before presenting to the user.

Scope:

Add reranker inference support via RerankerAdapter (already in Avocet model registry)
After initial job retrieval, rerank top-50 by (resume_text, listing_text) relevance score
Apply same reranking pass in the ATS rewriter: rerank candidate bullet phrasings by fit to job description
Wire to cf-core reranker service once available (see circuitforge-core ticket)
Free tier: local reranker (Qwen3-Reranker-0.6B or 1.5B); Paid: 8B variant

Out of scope: Training or fine-tuning the reranker on job-specific data (later milestone).

Acceptance criteria: Job listing feed order measurably improves on manual spot-check; ATS bullet suggestions ranked by JD fit score.

Related: Avocet model registry has RerankerAdapter + text-ranking pipeline tag support. Qwen3-Reranker-8B queued for download in Avocet.

**Context:** Embedding-based retrieval ranks job listings by cosine similarity to resume vectors, which misses nuance (a listing that mentions Python once scores similarly to one requiring 5 years of Python). A cross-encoder reranker pass over the top-50 results would significantly improve ranking quality before presenting to the user. **Scope:** - [ ] Add reranker inference support via `RerankerAdapter` (already in Avocet model registry) - [ ] After initial job retrieval, rerank top-50 by (resume_text, listing_text) relevance score - [ ] Apply same reranking pass in the ATS rewriter: rerank candidate bullet phrasings by fit to job description - [ ] Wire to cf-core reranker service once available (see circuitforge-core ticket) - [ ] Free tier: local reranker (Qwen3-Reranker-0.6B or 1.5B); Paid: 8B variant **Out of scope:** Training or fine-tuning the reranker on job-specific data (later milestone). **Acceptance criteria:** Job listing feed order measurably improves on manual spot-check; ATS bullet suggestions ranked by JD fit score. **Related:** Avocet model registry has `RerankerAdapter` + `text-ranking` pipeline tag support. Qwen3-Reranker-8B queued for download in Avocet.