SIM-RAG-Inspired Iterative RAG

Use this immutable Agentic Skill to run a bounded retrieve-generate-critique loop. It borrows SIM-RAG's inference architecture, but does not reproduce its Self-Practicing, Critic training, rationale generation, or experiments.

Contract

• Bind exactly one retriever, generator, and critic.
• Optionally bind one rewriter and one reranker.
• Keep the user's original query for reranking, generation, and critique.
• Treat rewritten queries as retrieval aids, never as evidence.
• Return an answer only when the Critic approves a direct, non-abstaining answer.
• Require evidence for every fact or reasoning hop needed by the answer.
• Regenerate once with the same evidence when a rejection is clearly about answer form only; ambiguous or evidence-related rejections continue retrieval.

Component selection

• Prefer BM25 for exact names, dates, titles, identifiers, and questions with strong lexical overlap with the corpus.
• Prefer Vector retrieval when paraphrases, synonyms, or semantic similarity are more important than exact token overlap.
• Use HyDE only with Vector retrieval, and only when a concise hypothetical passage is likely to bridge a vocabulary or intent gap. Never treat the HyDE passage as evidence.
• Use BGE Reranker when multi-hop reasoning, noisy candidates, or accumulated evidence make top-rank precision important.
• Leave optional slots empty when their expected retrieval benefit does not justify added model calls, latency, or local model cost.
• Always bind the required Generator and Critic after choosing the retrieval path.

Request options

• max_iterations defaults to 3 and must be positive.
• max_tokens controls answer generation only.
• critic_max_tokens controls critique generation, defaults to 4096, and must be positive.

Follow-up retrieval

Use at most three concrete missing-evidence issues from the Critic, bounded to 160 characters each. Use bounded feedback only when no usable issue exists. The trace exposes accumulated and newly added document IDs for every round.

Answer regeneration

Treat Critic feedback as answer-form-only when it clearly requests a shorter, more direct, or better-formatted answer and contains no missing-evidence signal. Regenerate once in that iteration using the original question, bounded Critic guidance, and the same accumulated evidence. Critique the revised answer again before returning it. Mixed or unrecognized rejection reasons must continue the normal retrieval loop.

Record a regeneration trace event with the original answer, bounded guidance, revised answer, and revised Critic result. An approved revision stops with critic_approved_after_regeneration.

Safe stopping

Stop with Insufficient evidence to answer reliably. when the iteration limit is reached or a rejected round contributes no new evidence. Treat generated abstentions as rejected even if a model Critic mistakenly approves them.