feature_request.yml

name: ✨ Feature request
description: Propose a new capability, provider, benchmark category, or architectural change.
title: "[feature] "
labels: ["enhancement", "triage"]
body:
  - type: markdown
    attributes:
      value: |
        Before filing, please skim [Discussions Ideas](https://github.com/Fearvox/Evensong/discussions/new?category=ideas)
        to see if this is better as an open-ended thread. File an issue when you have
        a concrete proposal with rough scope + success criteria.

  - type: textarea
    id: problem
    attributes:
      label: Problem / use case
      description: What are you trying to do that CCR / retrieval / benchmark can't currently do?
      placeholder: |
        I want to plug in a dense-vector stage 1 (BGE-M3 via BGE API or
        Cloudflare Workers AI) alongside BM25 in the hybrid pipeline, so
        semantic recall can rescue cases where BM25 term overlap sends
        the wrong doc to stage 2 (see Q44 Q56 Q59 in the 108-query run).
    validations:
      required: true

  - type: textarea
    id: proposal
    attributes:
      label: Proposed change
      description: Shape of the implementation. API sketch is ideal.
      placeholder: |
        New file: src/services/retrieval/providers/denseProvider.ts
        Factory: createDenseProvider({ endpoint, model }): VaultRetrievalProvider
        Composed via: createHybridProvider({ stage1: rrfFuse([bm25, dense]), stage2 })
      render: markdown
    validations:
      required: true

  - type: textarea
    id: alternatives
    attributes:
      label: Alternatives considered
      description: What other shapes did you reject, and why?
      placeholder: "Considered replacing BM25 entirely with dense — rejected because BM25 zero-cost is too good to give up."

  - type: dropdown
    id: scope
    attributes:
      label: Rough scope
      options:
        - S (≤ 1 file, ≤ 50 LOC)
        - M (new module, ≤ 300 LOC)
        - L (multi-module, new provider type, benchmark update)
        - XL (architectural — warrants an RFC discussion first)
    validations:
      required: true

  - type: textarea
    id: success
    attributes:
      label: Success criteria
      description: How will we know the feature works? Include a benchmark target if relevant.
      placeholder: |
        - Hybrid (BM25 + dense RRF → LLM) beats BM25-only Hybrid by ≥ 2pp top-1
          on the 108-query generated benchmark, 3 runs, stddev < 1pp
        - Latency overhead < 200ms p50 for dense stage
        - No regression on the 20-query manual benchmark
    validations:
      required: true

  - type: checkboxes
    id: willing
    attributes:
      label: Willing to contribute
      description: "Are you offering to PR this? No obligation either way."
      options:
        - label: I plan to open a PR for this.
        - label: I can test / review a PR but don't plan to write it.
        - label: I'm just flagging; no bandwidth to contribute.