# Review Retrieval Grounding

> Reviews retrieval and citations when an agent finds irrelevant evidence, misses sources, or makes unsupported claims; checks permissions and grounding separately from the broader agent threat model.

- Skill: `jovanipink/review-retrieval-grounding` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx get-skill add jovanipink/review-retrieval-grounding`
- Raw SKILL.md: https://api.skillmd.com/api/skills/jovanipink/review-retrieval-grounding/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: JovaniPink (https://skillmd.com/u/jovanipink)
- Updated: 2026-11-15
- Page: https://skillmd.com/skills/jovanipink/review-retrieval-grounding

---


# Retrieval and Grounding Quality Review

Review both retrieval quality and whether generated claims remain within the evidence retrieved. Read [the retrieval evidence matrix](#retrieval-evidence-matrix) when defining cases, metrics, or finding severity.

## Preconditions

Identify the intended users, tenants, questions, source authorities, corpus revision, ingestion path, retrieval configuration, model configuration, and decision the review must support. Do not use private records outside the authorized environment.

## Workflow

1. Map source ownership, licensing, authority, versions, effective time, ingestion, chunking, metadata, indexing, deletion, and refresh behavior.
2. Verify access control before retrieval. Test user, tenant, case, and document filters, including attempted cross-scope retrieval. Treat embeddings as sensitive derived data and evaluate whether inversion or reconstruction could expose source content.
3. Define representative, boundary, no-answer, conflicting-source, stale-source, adversarial, and known-failure queries. Pin the corpus and configuration.
4. Measure retrieval separately from generation. Assess relevant-source recall, irrelevant evidence, rank quality, coverage, freshness, latency, and missing-authority cases.
5. Map each material output claim to supporting evidence. Distinguish direct support, partial support, contradiction, unsupported inference, and missing evidence.
6. Review citation accuracy, source identity, units, dates, context, and whether a citation supports the nearby claim. Citation presence alone does not establish grounding.
7. Test prompt injection and retrieval poisoning in documents, metadata, tool results, and user content. Retrieved text remains untrusted data, not higher-priority instruction.
8. Define abstention, clarification, fallback, conflict disclosure, correction, and revocation behavior.
9. Measure whether retrieval changes intended product behavior, including correctness, relevance, safety, tone, and task usefulness. Record limitations, subgroup performance, cost, and what the cases cannot establish. Do not claim universal grounding from a sampled evaluation.

## Routing Boundaries

- Use `research-sources` for a person asking a current research question with cited primary sources.
- Use `verify-source-output` for byte-to-field or source-to-stored-output transformation evidence.
- Use `review-context-evidence` for a supplied context packet after retrieval.
- Use `design-agent-evaluations` to create the broader system evaluation contract.

## Output

Lead with the answer. Use the shortest format that retains the required evidence and limits; expand when requested or needed. The items below describe required content, not mandatory headings.

Return `Decision`, `Corpus authority`, `Access and isolation`, `Pinned configuration`, `Query set`, `Retrieval findings`, `Claim support findings`, `Citation findings`, `Poisoning and injection`, `Abstention and conflict behavior`, `Metrics`, `Limitations`, and `Required changes or tests`.

## Retrieval Evidence Matrix

Assess retrieval and claim support as separate stages.

| Stage | Questions |
| --- | --- |
| Source | Is the source authorized, licensed, current, versioned, and eligible for this user and purpose? |
| Ingestion | Were content, metadata, authority, time, access control, and deletion status preserved? |
| Embedding | Can stored or exposed vectors reveal protected source content through inversion, reconstruction, or similarity access? |
| Query | Does the case represent a real need, boundary, conflict, no-answer state, or attack? |
| Retrieval | Were relevant sources found, ranked, filtered, and returned within limits? |
| Claim | Is each material claim directly or partially supported, contradicted, or unsupported? |
| Citation | Does the cited source support the nearby claim with correct date, unit, and context? |
| Safety | Can document text, metadata, or a tool result change instructions, cross a tenant boundary, or trigger a tool? |
| Behavior | Does retrieval preserve the intended correctness, relevance, safety, tone, and usefulness of the application? |
| Recovery | Can the system abstain, ask, disclose conflict, correct a result, and revoke bad content? |

### Primary resources

- [OWASP vector and embedding weaknesses](https://genai.owasp.org/llmrisk/llm082025-vector-and-embedding-weaknesses/)
- [Google ADK safety and security](https://adk-labs.github.io/adk-docs/safety/)
- [NIST AI Resource Center](https://airc.nist.gov/)

A retrieved passage and a visible citation are evidence inputs. They are not proof that every generated claim is supported.

## Example

Synthetic example; not an observed client result.

Input: An answer cites a document that does not support its numerical claim.

Output: warning: retrieval succeeded but grounding failed. Remove or qualify the unsupported claim and trace whether a supporting source exists.

## Vector-risk source

The publisher-linked [official vector-risk data](https://genai.owasp.org/wp-json/wp/v2/llmrisk/254) identifies OWASP LLM08:2025 and the same canonical article. Use it as risk guidance alongside the reviewed system's concrete authorization and retrieval evidence; it does not establish that controls or isolation were observed.

