/flashrag-evidence

name: flashrag-evidence
description: "Local evidence retrieval (FlashRAG-style) for VCO/vibe: search protocols/config/skills docs and return citeable snippets with file+line anchors."

FlashRAG Evidence (VCO)

When to use

Use this skill when you need **grounded, citeable evidence** from local documentation/configuration to support VCO decisions or recommendations, especially for:

• VCO routing / pack selection rationale
• Protocol compliance (think/do/review/team/retro)
• Config semantics (thresholds, overlays, governance)
• “Show me where this rule comes from” / “give me the exact snippet”

This skill is **not** a replacement for GitNexus (code dependency graph) or web search. It focuses on **local docs and config**.

Inputs

• Query: what you’re trying to verify (short, concrete)
• Optional: corpus root(s) to search (defaults below)

Default corpus (evidence plane)

1. VCO core docs/config inside `~/.codex/skills/vibe/`:

• `protocols/`, `config/`, `references/`, `scripts/router/`

2. Skills catalog (`~/.codex/skills/**/SKILL.md`) for tool capability evidence 3. (Optional) Project-local VCO overlays under the current workspace, if present

Workflow (Lite, no heavy deps)

1. Run the evidence retriever script:

• Windows PowerShell:
• `python C:\Users\羽裳\.codex\skills\flashrag-evidence\scripts\flashrag_evidence.py --query "…" --topk 8`

2. (Optional) Enable a faster FlashRAG-style BM25 backend (`bm25s`)

• Preflight (checks vendoring + env; does NOT read secrets):
• `pwsh C:\Users\羽裳\.codex\skills\vibe\scripts\ruc-nlpir\preflight.ps1`
• Manually create an isolated venv for the vendored runtime and install only the minimal packages you need. The old `install-upstreams.ps1` auto-install path has been removed on purpose.
• Use bm25s engine:
• `C:\Users\羽裳\.codex\_external\ruc-nlpir\.venv\Scripts\python.exe C:\Users\羽裳\.codex\skills\flashrag-evidence\scripts\flashrag_evidence.py --engine bm25s --query "…" --topk 8`

3. Use the returned snippets as P5 evidence:

• **[Command]** the exact command you ran
• **[Output]** the top snippets (path + line anchor)
• **[Claim]** the conclusion you draw (only what the evidence supports)

4. If coverage is low:

• Expand `--roots` to include the project workspace
• Increase `--topk`
• Fallback: targeted `rg -n` on the most likely file(s)

Outputs

The script prints ranked evidence items:

• `path` + `line` (1-based) for quick navigation
• `score` for ranking
• `snippet` (short, safe to quote)

Notes (non-redundancy)

• If you need **code call chains / blast radius**, use GitNexus overlays (not this).
• If you need **latest web facts**, use web search / deep research tools (not this).