Definition
Sub-document retrieval is the practice of splitting pages into smaller units, such as sections, passages, table rows or list items, and indexing, retrieving and ranking those units individually instead of treating each page as one indivisible document. In an AI search pipeline, what reaches the language model is usually a handful of passages that happen to live on various URLs, not the URLs themselves.
Perplexity described this explicitly when it launched its Search API in September 2025. Its index 'tracks over 200 billion unique URLs', and its parser decomposes documents 'into self-contained spans, each of which can be individually retrieved and ranked at query time'. Results are scored 'at both the document and sub-document levels', and the pipeline is 'designed to surface the most atomic units possible to the model' so the model is not fed irrelevant text from elsewhere on the page (Perplexity). Most retrieval-augmented generation systems work the same way, splitting documents into chunks of a few hundred tokens and embedding each chunk separately.
The reason is practical. A model's context is limited and expensive, and, as Perplexity's engineers note in the same post, a large body of prior work shows model performance degrading as context grows and fills with irrelevant text. Sending the best five passages from five sources beats sending one full page.
This changes the unit of optimization. A page can rank well in classic search on the strength of the whole document while containing no passage that works on its own, because every useful sentence relies on context established paragraphs earlier. Anthropic's write-up on Contextual Retrieval gives the textbook case: a chunk reading 'The company's revenue grew by 3% over the previous quarter' is useless for a question about ACME's Q2 2023 results, because the company name and period appear elsewhere in the filing (Anthropic). Engineers fix that by adding context to chunks before indexing. Writers can fix it at the source:
- Restate the subject in each section instead of relying on 'it' or 'the product'.
- Put the answer in the first sentence or two of the section, not after a long lead-in.
- Keep a claim and its supporting figure or source in the same paragraph.
- Use headings that name the question or entity the section covers.
- Give tables descriptive headers and full row labels so a single row still makes sense.
These are the mechanics behind content chunking and answer-ready content. Sub-document retrieval is related to Google's passage ranking, but the two differ: Google said its 2021 passage ranking system still indexes whole pages and uses passages as an additional ranking signal, while AI retrieval pipelines such as Perplexity's index and rank the passages themselves.
It also explains a common measurement puzzle. An engine may quote one sentence from deep inside a long guide and ignore its introduction, or cite a page that ranks nowhere for the head term. When AI crawler logs show heavy fetching of a page that rarely earns citations, check whether its passages stand on their own before rewriting anything else.
Examples of Sub-Document Retrieval
- Perplexity cites a single definition paragraph from a 4,000-word guide on employee onboarding and none of the rest of the page.
- A writer changes 'It integrates with Salesforce in under an hour' to 'Acme Sync integrates with Salesforce in under an hour', so the sentence still identifies the product when retrieved alone.
- A comparison table row reading 'Plan: Team, Price: $12 per user per month, Seats: up to 50' is quoted directly in an AI answer because the row carries all the context it needs.
- Crawler logs show OAI-SearchBot fetching a long guide daily, yet ChatGPT never cites it; the guide puts every answer below three paragraphs of scene-setting, and moving the answers up leads to citations within weeks.
