RupuContext

Lint the pack. Don't pay twice.

Part of the Rupu family.

Local CLI. Deterministic JSON audit reports. Policy gates for CI. Technical signals — not certification, not token pricing, not another LLM call to "evaluate."

Install

Requires Python 3.9+.

pip install rupucontext
rupucontext --help

Quick start

rupucontext scan fixtures/dup-pack.jsonl
rupucontext scan fixtures/dup-pack.jsonl --fail-on-overlap
rupucontext compare fixtures/corpus.jsonl fixtures/questions.jsonl

Policy gates exit 2 when thresholds are exceeded. Exit 1 is reserved for errors. The JSON audit report (including gate) is written either way.

Commands

scan reports retrieve duplicates in exact_duplicates / near_duplicates, and cross-role overlap in cross_segment (e.g. policy pasted in system and retrieve). compare uses the same segment_text unit as scan.

A JSONL file with one or more pack_id values always emits result.results[] (one element per pack).

Gates are flags, not subcommands:

duplicate_rate and record_rate use retrieve segments only as the denominator (rate_denominator: retrieve_segments). --near-duplicate-threshold (default 0.85) controls detection of near-duplicates. Policy gates use rates, not the Jaccard cutoff.

Input format (JSONL)

One line per segment. Group by pack_id:

{"pack_id": "req-001", "role": "system", "text": "You are a helpful assistant..."}
{"pack_id": "req-001", "role": "retrieve", "chunk_id": "c42", "text": "Refund policy: items within 30 days..."}
{"pack_id": "req-001", "role": "retrieve", "chunk_id": "c17", "text": "Refund policy: items within 30 days..."}
{"pack_id": "req-001", "role": "user", "text": "Can I return this?"}

Roles: system, retrieve, history, user (extensible). chunk_id optional but recommended for retrieve segments.

Audit report

Reports follow: input → configuration → method → result → (optional) gate.

{
  "tool": "rupucontext",
  "version": "0.1.0",
  "command": "scan",
  "input": {
    "path": "fixtures/policy-twice.jsonl",
    "segments": 3,
    "packs": 1
  },
  "configuration": {
    "near_duplicate_threshold": 0.85,
    "fail_on_overlap": true
  },
  "method": {
    "unit": "segment_text",
    "exact": "text_exact_v1",
    "normalized": "text_normalized_v1",
    "near": "jaccard_char_shingles_v1"
  },
  "result": {
    "results": [
      {
        "pack_id": "req-policy",
        "segment_count": 3,
        "retrieve_segment_count": 1,
        "exact_duplicates": {
          "pairs": 0,
          "segments_flagged": 0,
          "duplicate_bytes": 0,
          "duplicate_rate": 0.0,
          "rate_denominator": "retrieve_segments",
          "evidence": []
        },
        "near_duplicates": {
          "pairs": 0,
          "segments_flagged": 0,
          "record_rate": 0.0,
          "rate_denominator": "retrieve_segments",
          "evidence": []
        },
        "cross_segment": [
          {"from": "retrieve", "to": "system", "method": "exact", "overlap": 1.0}
        ]
      }
    ]
  },
  "gate": {
    "passed": false,
    "rules": [
      {"metric": "overlap_pairs", "actual": 1, "threshold": 0, "passed": false}
    ]
  }
}

fixtures/dup-pack.jsonl (quick start) flags retrieve∩retrieve duplicates; cross_segment may be [] when no cross-role overlap is detected.

Byte counts and overlap ratios — not token estimates. Contract details: docs/AUDIT.md.

Exit codes

# .github/workflows/rupucontext.yml
- run: rupucontext scan fixtures/clean-pack.jsonl --fail-on-overlap

What it is not

• Not token pricing (no tiktoken, no USD)
• Not prompt versioning or observability proxy
• Not semantic / paraphrase matching (on purpose)

If you can't export a pack or a trace, you're not the user yet.

Development

git clone https://github.com/EmanuelCorreaAR/rupucontext.git
cd rupucontext
pip install -e ".[dev]"
pytest

Status

0.1.0 — scan + compare; policy gate flags; deterministic audit JSON.

Next: stabilize audit contract toward 1.0.

License

Apache License 2.0