Local Evidence Finder

Turn a project question into local, citable evidence. Keep private files on the machine and treat retrieved passages as evidence, not as a complete answer by themselves.

Workflow

Resolve the narrowest directory that contains the relevant project. Do not default to a home directory or drive root.

Run a semantic search with the user's wording:

& "<skill-dir>/scripts/run.ps1" search --root "<project-dir>" --query "<question>" --top-k 5 --format json

Inspect results. Prefer passages with direct wording and useful context; do not hide a low-confidence result.
Open the cited files when a claim requires more surrounding context. Preserve the returned line ranges when reporting evidence.
Answer from the evidence and distinguish facts, conflicts, and inferences. Say when the local corpus does not support a conclusion.

Operations

Search and update a stale index automatically:

& "<skill-dir>/scripts/run.ps1" search --root "<project-dir>" --query "<question>" --top-k 5

Force a rebuild after large file changes:

& "<skill-dir>/scripts/run.ps1" reindex --root "<project-dir>"

Check the resident service and execution device:
```
& "<skill-dir>/scripts/run.ps1" status
```

Stop the local service:

& "<skill-dir>/scripts/run.ps1" shutdown

Read references/cli.md only when exact arguments, result fields, supported file types, or exit codes are needed.

Evidence Rules

Cite the returned relative path and line range beside each factual claim.
Quote only the minimum excerpt needed and never invent missing context.
Treat low_confidence: true as a retrieval warning. Refine the query or inspect the project manually before making a strong claim.
If files disagree, show both passages and describe the conflict.
Never send file contents to a remote model or service as part of this skill's retrieval path.
Do not index outside the user-authorized project root.

Setup Boundary

The first run creates an isolated runtime under %LOCALAPPDATA%/local-evidence-finder, installs pinned dependencies, verifies a published BGE model bundle, and converts its quantized ONNX export to OpenVINO IR. Network access is required only for that setup. Later searches run on localhost and can work offline. Read references/model-provenance.md when reviewing licensing, checksums, or model origin.