liveBook · Manning

Self-correction

As explained in Retrieval Augmented Generation, The Foundational Ideas

Self-monitoring and correction is a system capability that evaluates outputs and information quality and initiates corrective actions when problems are detected.

Self-correction is the capability to evaluate generated outputs and retrieved information, detect quality gaps, and trigger corrective actions before an answer reaches the user. In retrieval-augmented generation, it turns a one-shot pipeline into a feedback loop: assess the evidence or draft, then retrieve again, revise the plan, or continue searching when support is insufficient. section 1.5.4 section 11.1

A self-correcting system can:

  • Score confidence in an answer or intermediate result. section 1.5.4
  • Verify factual claims and detect missing information. section 1.5.4
  • Decide whether to retrieve, what to retrieve, and whether another retrieval cycle is needed. section 11.1
  • Loop back to planning or retrieval instead of returning a silently failed or hallucinated answer. section 11.1

Financial compliance assistant

A compliance assistant retrieves evidence for a regulatory question, checks whether the answer covers all required parts, and searches again when evidence is incomplete, stale, poorly ranked, or insufficiently specific. This makes corrective retrieval a response to identifiable production failure modes rather than an assumption that more complexity is always better. section 11.1

Figure 11.1 From Passive Pipelines to Agentic Loops: The shift from linear execution to cyclical evaluation.

Use the diagram to compare passive one-query/one-answer execution with cyclical evaluation and correction. section 11.1

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Explains confidence scoring, factual verification, gap detection, and iterative validation, but does not specify implementation algorithms.
Examples
Uses a financial compliance assistant, research agents, confidence scoring, factual verification, and gap detection as concrete examples.
What you can do after
Evaluate/understand when an answer should trigger validation, renewed retrieval, or a return to planning.
How it connects
Connects self-correction with agentic RAG, dynamic retrieval, query formulation, planning, factual verification, and multi-step reasoning.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Retrieval Augmented Generation, The Foundational Ideas. The section links point to the parts of the book it draws on.