recallA retrieval
field guide

Emerging Retrieval Paradigms

Build this skill

Describe tree-based, vectorless and other emerging retrieval approaches and evaluate new methods critically.

Retrieval is still changing fast

The basic recipe of chunking documents, embedding chunks and searching by similarity has dominated RAG since around 2020. Researchers and engineers keep proposing alternatives, mostly aimed at its weak spots: chunks that lose structure, embeddings that blur details, and single searches that cannot reason. Some of these ideas are already in production; others are promising experiments. This topic surveys a few directions worth knowing, with an eye on what problem each one tries to solve.

Retrieval over document trees

Long documents have structure: a table of contents, chapters, sections, subsections. Flat chunking throws that structure away. Several approaches keep it and search along it.

RAPTOR (2024) builds a tree of summaries from the bottom up: it clusters chunks, writes a summary of each cluster, clusters those summaries, and so on, until it has a tree from fine details to broad themes. A question can then match at whatever level of detail it needs, from a specific clause to a whole-document summary.

Reasoning-based tree navigation works top-down instead. The language model is given a document's table of contents (or a generated outline), reads it, decides which section is most likely to contain the answer, opens that section, and repeats, much as a person uses a long manual. Open-source projects such as PageIndex follow this "vectorless" approach, using the model's reasoning over document structure instead of embedding similarity.

Academic Regulationstable of contentsCh. 2 AttendancechosenCh. 4 Leavemodel chooses thischosen4.1 Medical4.2 Duty leaveCh. 6 Exams

Instead of comparing embeddings, the model reads the outline and decides where to look, step by step.

Vectorless retrieval

"Vectorless" is a loose label for approaches that skip embeddings entirely. Tree navigation is one kind. Another is simply keyword search plus a strong model: with good BM25 and a language model that rewrites queries and reads results carefully, some teams find they do not need a vector database at all for their collection. A third is having an agent explore files directly, listing folders, searching text with tools like grep and opening files, the way coding assistants search a code repository.

The appeal is simplicity and transparency: no embedding model to choose or refresh, and every step is readable. The costs are more model calls per question (each navigation step is a model call), slower answers, and a dependence on documents having a meaningful structure.

The program below shows tree navigation in miniature. A stand-in "model" picks the section whose title shares the most words with the question, then descends.

python · runs in your browser

Expected output:

text
Academic Regulations > 4 Leave > 4.1 Medical leave
answer text: Up to 10 days per semester; more needs the Dean.

choose plays the part of the model: it reads the titles at the current level and picks the one sharing the most words with the question. The while loop keeps descending as long as the current node is a dictionary, meaning it still has sub-sections, and stops when it reaches text. No embeddings are involved at any point.

Other directions to watch

Direction The problem it targets Idea
Learned sparse retrieval Keywords miss synonyms Neural models produce keyword weights, including related words
Multi-vector retrieval Single vectors blur details One vector per token, compared with late interaction
Visual document retrieval Parsing loses layout and charts Embed images of pages directly and search those
Graph-based retrieval Multi-hop and whole-collection questions Store entities and relations; summarise communities
Agentic retrieval One search is not enough The model plans and repeats searches as needed
Long-context and cached context Retrieval misses evidence Read far more, or everything, at once

Reading research with a critical eye

New retrieval methods appear every month, often with impressive benchmark numbers. When you read about one, a few questions keep you grounded:

  • What problem does it solve, and do you have that problem? A method for multi-hop questions is wasted on a collection of FAQs.
  • What does it cost? Count the extra model calls, indexing time and storage.
  • What was it compared against? A fair comparison includes strong baselines such as hybrid search with reranking, not just plain vector search.
  • Does it work on your data? Reproduce a small version on your own test questions before adopting it.

Follow your curiosity.

↑ ↓ to explore · Enter to open · Escape to close