Skip to main content
Back to timeline
CloudflareSource publication:

AI Search now writes file summaries during indexing, pushing topics into chunk context headers while LLM costs bill separately

Synopsis

Cloudflare's AI Search can now generate a summary, topics and keywords for a file while it indexes it, for each text document that produces at least three chunks. The summary and keywords go into the item's context metadata, and the topics are added to the context header of that file's chunks to help retrieval. The feature is off by default, turning it on or changing the setting triggers a full reindex, and LLM calls are billed as Workers AI usage rather than against the AI Search ingestion allotment.

Source-provided article image: Changelog
Cloudflare

Interpretation

AI Search's indexing pipeline gains a step: while splitting text, it calls a large language model to produce a summary, topics and keywords for every text document that yields at least three chunks. The summary and keywords are stored in the item's context metadata, and the topics enter the context header of that file's chunks as extra information available at retrieval. The behavior is controlled by the indexing option indexing_options.file_summary, which is off by default and can be enabled in the dashboard or by setting it to true when creating or updating an instance.

The generated topics are placed in the context header of the file's chunks, while the summary and keywords stay in item-level context metadata. An acontext value already set on the item is kept rather than overwritten. The scope is limited to text documents that produce at least three chunks; documents below that chunk count do not enter the summarization step.

Summary calls go through the instance's AI Gateway, or the default gateway if none is set, and are billed to the account as Workers AI usage. These costs are not part of the AI Search ingestion allotment and must be covered separately. Each new or modified file is summarized once, so cost grows with the amount of content indexed; changing the setting triggers a full reindex, and indexing_options is replaced as a whole on every update, so omitted options return to their defaults.

Perspective

The note covers AI Search indexing of text documents and only those producing at least three chunks; enabling, billing attribution and reindex behavior are all bounded by this product setting.

No comparison data is given for summary quality or retrieval benefit; each file is summarized once and regenerated only after modification; indexing_options is replaced as a whole, so omitted options return to defaults.