RAG

OpenRAG-Skill

Skill invocation remains $rag.

Use this skill only when the relevant material is already in context:

• The user pasted one or more documents into the chat.
• The conversation already contains policies, manuals, contracts, specs, or FAQs.
• The task requires a cited answer, extraction, comparison, or audit grounded only in that material.

Do not use this skill for:

• Web search or external knowledge retrieval
• Large corpora that are not already present in context
• Embeddings, vector search, BM25, indexing, or reranking
• Repository-wide code navigation as the primary task

Input Modes

Raw Context Mode

Use this when the user provides unstructured source material.

[CONTEXT]
Source A:
...

Source B:
...

[TASK]
What should be answered or extracted?

[OUTPUT_MODE]
answer | extract | compare | audit

First package the material into Evidence Units, then answer from those units.

Packaged Context Mode

Use this when the user already provides labeled evidence.

[E1] Source: ...
Section: ...
Content: ...

[E2] Source: ...
Section: ...
Content: ...

Skip packaging and move directly to localization, reconciliation, and answering.

Evidence Unit Format

An Evidence Unit is a citation container, not a retrieval chunk.

Required fields:

• ID (E1, E2, ...)
• Source
• Section
• Content
• Optional Notes for date, scope, version, or jurisdiction

Core Protocol

Execute these stages in order.

Stage 1: Context Packaging

Purpose: turn raw material into stable, referenceable Evidence Units.

Rules:

• Preserve source boundaries.
• Preserve headings, numbering, and clause labels.
• Split only on visible structure such as headings, bullets, numbered clauses, or paragraph breaks.
• For long documents, package hierarchically: source -> section -> clause.

Deliverable:

• Evidence Register: the numbered evidence list (E1, E2, E3, ...)

Stage 2: Query Decomposition and Coverage Planning

Purpose: define what must be proven before answering.

Rules:

• State the core question.
• Break it into atomic sub-questions.
• Classify the task as fact, rule, comparison, extraction, or audit.
• Mark which sub-questions are mandatory for a complete answer.

Deliverable:

• Coverage Table: each sub-question mapped to the evidence needed to support it

Stage 3: Evidence Localization and Extraction

Purpose: identify the smallest sufficient evidence set.

Rules:

• Prefer direct statements over inference.
• Track unresolved sub-questions separately instead of guessing.
• Preserve exact wording for policy, contract, and compliance tasks.
• If support is indirect, treat it as lower-confidence support.

Deliverable:

• Evidence Ledger: direct support, indirect support, and unresolved gaps for each sub-question

Stage 4: Cross-Check and Reconciliation

Purpose: prevent unsupported synthesis and hidden contradictions.

Rules:

• Check whether evidence units agree, narrow, or conflict.
• Track normative force for each controlling claim (SHALL, SHALL NOT, SHOULD, SHOULD NOT, MAY).
• Prefer more specific evidence over summaries.
• Prefer time-bounded or versioned evidence when recency is explicit in the source.
• Never merge claims with different normative force as if they were equivalent.
• Do not flatten unresolved conflicts into one rule.

Deliverable:

• Conflict Ledger: scope, date, version, exception, and contradiction notes used in the final response

Final Sweep

Before drafting, run one explicit omission check:

• Scan for exception terms such as except, unless, however, only if, and notwithstanding.
• Scan for dates, effective times, and version labels.
• Scan for scope boundaries such as role, region, product tier, or approval authority.
• Scan for wording that is legally or operationally sensitive, especially modal force.

If the sweep finds a missing constraint, add it to the ledgers before answering.

Stage 5: Cited Answer Generation

Purpose: produce a stable, reviewable output.

Rules:

• Every material claim must be citation-backed.
• Unsupported claims must be omitted.
• Missing support triggers a partial answer or a refusal.
• Do not add outside knowledge unless the user explicitly asks for reasoning beyond the provided material.
• If a detail is absent, say not specified in the provided evidence instead of inferring from terminology, category, or common practice.
• Return the answer directly in the response; do not create helper files or external artifacts unless the user explicitly asks.

Deliverable:

• A contract-formatted answer with aligned evidence IDs

Normative Force Lock

When the source uses modal or normative language:

• Preserve SHALL, SHALL NOT, SHOULD, SHOULD NOT, and MAY exactly in meaning.
• Do not rewrite SHOULD or MAY as if they were mandatory.
• Do not rewrite SHALL NOT or SHOULD NOT as softer preference language.
• If two sources discuss the same topic with different force, surface that as a difference or conflict.
• In normative comparisons, name the force explicitly inside each differing requirement.

Internal Work Products Stay Internal

Evidence Register, Coverage Table, Evidence Ledger, and Conflict Ledger are internal work products by default.

Do not emit them as extra top-level sections unless the user explicitly asks to see the intermediate reasoning artifacts. Do not emit TODO lists, planning notes, or scratch reasoning.

Output Modes

OUTPUT_MODE changes the structure inside the Answer section.

answer

Use for direct conclusions.

• Return the shortest sufficient conclusion set.
• Keep each conclusion explicitly cited.

extract

Use for field extraction or direct pull-through.

• Return field names, extracted values, and not found where evidence is missing.
• Avoid narrative unless the task explicitly asks for it.

compare

Use for side-by-side source comparison.

• Organize the Answer section as Shared Ground, Differences, Conflicts, then Bottom Line.
• Put a point in Shared Ground only when both scope and normative force match across the cited evidence.
• If one source says SHALL and another says SHOULD or MAY, that belongs in Differences, not Shared Ground.
• If force differs and the overlap is only topical, prefer Shared Ground = None. instead of a vague common denominator.
• Never use words like require, must, or prohibit in Shared Ground unless all cited evidence support the same mandatory force.
• In normative text, each bullet in Differences should name the force explicitly, such as Passwords [SHALL]: ....
• Bottom Line may restate only already-cited differences; do not introduce new rankings, causes, or unstated rationale.
• Keep scope attached to the source that owns it.

audit

Use for policy, controls, and compliance-style reviews.

• Organize the Answer section as Requirements, Exceptions, Gaps, then Risks when relevant.
• Do not omit scattered constraints that narrow the main rule.

Response Contract

Use these top-level sections in this exact order:

1. Answer
2. Evidence Map
3. Conflicts or Gaps
4. Need More Material (only when partial or refused)

Rules:

• Use no extra top-level sections before Answer.
• Every non-trivial claim in Answer must end with one or more evidence IDs such as [E3] or [E2][E5].
• Evidence Map must map each major conclusion to the exact evidence IDs that support it.
• If a claim cannot be supported, it must not appear in Answer.
• Do not emit Need More Material when the answer is complete.
• Never emit placeholder text such as None. inside Need More Material.
• Keep the wording stable and easy to score.

Refusal Contract

If the provided material does not support the core question:

• Set Answer to INSUFFICIENT_EVIDENCE.
• State exactly what is missing.
• List the additional material needed.
• Do not provide a speculative fallback.

Conflict Contract

When evidence conflicts:

• Cite the conflicting evidence IDs directly.
• Explain whether the conflict is caused by scope, date, version, or an irreconcilable contradiction.
• Use conditional conclusions instead of a flattened summary.

Preferred patterns:

• If E2 governs, then ...
• If E5 is the current version, then ...
• The sources conflict and do not support a single unconditional conclusion.

Operating Defaults

• Optimize for long-form documents first.
• Prefer complete traceability over broad coverage.
• Use exact wording when a small wording change would alter the meaning.
• If only part of a multi-part request is supported, answer only that supported part and mark the rest as a gap.

References

Use these bundled references as needed:

• references/protocol.md for the detailed operating procedure
• references/templates.md for copy-paste templates
• references/failure-modes.md for failure handling
• references/methodology.md for positioning and limits

See the examples/ and evals/ folders for worked examples and test cases.