← Back to concept explainers

Concept Fable

Recursive retrieval

Conclusion first:Recursive Retrieval is a retrieval approach in RAG or knowledge retrieval that generates follow-up queries based on already retrieved content, tracks new entities or fills in missing evidence. It suits multi-hop questions, complex fact-checking and long-chain business analysis. The key is to control recursion depth, deduplication, stopping conditions and evidence records, to avoid infinite expansion or drifting further off target the more you search.

Recursive Retrieval · Retrieval

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

Three shelves of books pulled from a single index card in the archive

The national history office received a question: why did that salt tax reform thirty years ago fail? The new clerk first searched by "salt tax reform" and brought back three volumes of edicts. After reading them, he knew only the outcome, not the cause.

The curator told him to continue tracing clues from the volume. The decree mentioned that "shipping in the southeast was blocked," so he went to the shipping cabinet; in the shipping volume appeared "private port commissions," which led him to the coastal defense case; in the coastal defense case there was a memorial from a local official, mentioning that rising grain prices affected salt prices.

The scribe complained it was like digging a well, one layer after another. The curator said the answer to this kind of question was never in the first shovel of soil. The material found the first time should help you generate a more precise question for the next round.

They also set stop-loss rules. Each round of pursuit had to explain why to continue, and any new volume found had to be able to change the answer; if after three rounds only side branches remained, they stopped and organized. Otherwise the archive would be dragged into a bottomless pit by clues.

The final report linked the imperial edict, shipping, private ports, grain prices, and local memorials into one chain. If only the first round were investigated, the conclusion would stop at 'the court's orders were ineffective'; after progressive investigation, it became clear that multiple factors affected one another.

The new scribe thought this was too slow and wanted to pull out all the cabinets at once. The curator stopped him: moving everything would overwhelm, moving too little would break the thread. The real method is to read one thread end, then follow it to the next cabinet; each time a new scroll is found, judge whether it brings a new clue.

Only after several rounds did the pattern of the salt tax failure surface: blocked shipping pushed up private port commissions, private port commissions changed grain prices, and grain prices in turn made the salt price policy fail. The answer was not opened by the first key; behind each door another key appeared. Complex problems often require reading and chasing as you go, until the clues no longer extend. In the end the new clerk compiled a clue map rather than a pile of unrelated files. The curator said that when pursuing a deep problem, each piece of material found should change the next search; if the direction of search is never rewritten by new evidence, that is just repeatedly rummaging through cabinets.

The scribe understood that some searches aren't about grabbing everything at once, but about reading while asking and asking while searching. Each round of material should open the next, more relevant door, until the question is sufficiently explained—not buried under a pile of paper.

Reveal

This story is about: recursive retrieval

Recursive Retrieval is a retrieval approach in RAG or knowledge retrieval that generates follow-up queries based on already retrieved content, tracks new entities or fills in missing evidence. It suits multi-hop questions, complex fact-checking and long-chain business analysis. The key is to control recursion depth, deduplication, stopping conditions and evidence records, to avoid infinite expansion or drifting further off target the more you search.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • One index lookup: ordinary single-turn retrieval
  • Chamber of commerce dispute: a complex problem requiring multi-hop evidence
  • Take a few more scrolls: blindly expanding top-k
  • New names, ship numbers, former names: clues for the next round of queries generated from search results
  • Follow-up questions and stopping reasons: recursion depth, stopping conditions, and traceability
  • Retrieval Room: A Recursive Retrieval Process for Complex Tasks

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

multi-hop retrievalquery expansion
京ICP备2026014272号