Product Kit · 1.0.0

RAG Evidence Preparation Kit

Turn already-retrieved records into bounded citation-ready evidence with exact quotes and explicit omissions.

RAG and research pipeline builders who need deterministic evidence preparation without another relevance or model layer.

Submitted-score ordering is not reranking or relevance scoring. Excerpt artifacts contain exact caller text and must be redacted before sharing.

The sample was executed through the real deterministic preparation core during this build.

Preparation envelope

{
  "ok": true,
  "schema_version": "utilito.product-kit.v1",
  "kit": {
    "id": "rag-evidence-preparation-kit",
    "version": "1.0.0"
  },
  "result": {
    "status": "review_required",
    "evidence": [
      {
        "source_id": "guide-a",
        "quote": "Use exact retrieved text as citation evidence.",
        "submitted_score": 0.92,
        "url": "https://example.test/guide-a",
        "title": "Guide A",
        "id": "evidence-337bf7c77ae7fe4c"
      },
      {
        "source_id": "guide-b",
        "quote": "Context budgets retain complete quotes only.",
        "submitted_score": 0.81,
        "title": "Guide B",
        "id": "evidence-3cca4b25a4769b16"
      }
    ],
    "ordering": {
      "policy": "submitted_score_desc",
      "score_meaning": "caller_submitted_not_reranked_or_verified"
    },
    "dedupe": {
      "policy": "remove",
      "identity": "exact canonical JSON [source_id,quote]; metadata differences do not change citation identity",
      "duplicate_occurrences": 0,
      "removed": 0,
      "retained": 0
    },
    "coverage": {
      "input_evidence_count": 2,
      "valid_evidence_count": 2,
      "ordered_evidence_count": 2,
      "retained_evidence_count": 2,
      "omitted_evidence_count": 0,
      "duplicate_omitted_count": 0,
      "context_budget_omitted_count": 0,
      "input_quote_bytes": 90,
      "retained_quote_bytes": 90,
      "omitted_quote_bytes": 0,
      "quote_byte_coverage": {
        "numerator": 90,
        "denominator": 90
      }
    },
    "omissions": [],
    "context_pack": null,
    "context_summary": {
      "schema_version": "utilito.rag-evidence-context-summary.v1",
      "evidence_count": 2,
      "retained_evidence_count": 2,
      "context_budget_bytes": 8192,
      "raw_evidence_included": false
    },
    "report": {
      "format": "markdown",
      "mime_type": "text/markdown",
      "contains_raw_evidence": false,
      "warning": "Raw evidence omitted in metadata-only mode.",
      "content": "# RAG Evidence Preparation report\n\n> Deterministic evidence report generated without a model call.\n\n- Recipe: `rag\\-evidence\\-pack@1\\.0\\.0`\n- Source contract: `utilito.context-pack.v1`\n- Storage: response-only; not stored by Artifact Ops\n\n## Transformation receipt\n\n| Metric | Value |\n|---|---:|\n| Input bytes | 346 |\n| Output bytes | 346 |\n| Estimated tokens before | 87 |\n| Estimated tokens after | 87 |\n| Input records | 2 |\n| Output records | 2 |\n| Omitted records | 0 |\n| Omitted fields | 0 |\n| Deduplicated | 0 |\n| Truncated | no |\n\n## Evidence\n\nMetadata-only mode: raw evidence IDs, source IDs, titles, URLs, quotes, and model-ready output are omitted. Retained evidence items: 2.\n\nRaw excerpts omitted in metadata-only mode.\n\n"
    },
    "safe_actions": [
      {
        "code": "REVIEW_EVIDENCE_ORDER",
        "message": "Review caller-submitted score ordering or preserved input order before using the evidence.",
        "automatic": false
      },
      {
        "code": "KEEP_RETRIEVAL_SOURCE",
        "message": "Keep the original retrieval response as the authoritative source for omitted records and submitted metadata.",
        "automatic": false
      }
    ],
    "semantic_relevance_inferred": false,
    "reranked": false,
    "summarized": false,
    "fetched": false,
    "embedded": false,
    "submitted_content_executed": false
  },
  "receipt": {
    "input_bytes": 536,
    "output_bytes": 2639,
    "operations": 4,
    "omitted_items": 0,
    "warnings": [],
    "input_records": 2,
    "retained_records": 2,
    "duplicate_occurrences": 0,
    "duplicates_removed": 0,
    "duplicates_retained": 0,
    "context_budget_omissions": 0,
    "input_quote_bytes": 90,
    "retained_quote_bytes": 90,
    "omitted_quote_bytes": 0
  },
  "provenance": {
    "deterministic": true,
    "model_used": false,
    "network_used": false,
    "storage_used": false,
    "core_versions": {
      "rag_evidence_preparation_kit": "1.0.0",
      "context_pack": "1.0.0",
      "evidence_report": "1.0.0"
    }
  },
  "error": null
}

Free runnable capability

Prepare one bounded retrieval response with citation IDs, exact quote evidence, Context Pack, receipts, and report with no signup.

Builder capability

The existing Builder Pilot adds higher bounded volume and batch preparation; it does not add fetching, embeddings, semantic ranking, model calls, storage, or telemetry.

See the existing Builder Pilot

Response-only privacy

Submitted retrieval records are sent only to the response-only API when you select Prepare evidence; they are not fetched, stored, embedded, reranked, summarized, modeled, or logged as payload telemetry.

Limits and trust boundary

  • Hosted serialized input is limited to 32,768 UTF-8 bytes, 500 records, and exact quotes of at most 8,192 UTF-8 bytes each within the shared request limit.
  • submitted_score_desc uses only finite caller-mapped numeric scores and stable input-index ties. It is submitted-score ordering, not Utilito reranking, relevance scoring, or verification.
  • Exact duplicate identity is canonical [source_id, quote]; title, URL, and score differences do not change citation identity. Duplicate sources with different quotes remain distinct.
  • Evidence IDs are deterministic citation labels, not cryptographic integrity or security hashes. Kept exact duplicates use occurrence suffixes, so inserting or removing an earlier identical occurrence can change later duplicate labels.
  • Metadata mode removes raw source IDs, evidence IDs, URLs, titles, and quotes from context/report artifacts. Excerpt artifacts contain exact submitted evidence and must be redacted before sharing.