I'm Nitin Kumar, Lead Project Coordinator at Debugged Pro. I work with software and IT agencies in India and the USA that need clear vendor facts before they put RAG or agent search into a client sprint. Cloudflare AI Search moved from beta talk to general availability on October 1, 2026.

Cloudflare's product blog says AI Search is generally available with native image embeddings, OCR for PDFs, and larger file support for several formats. The same post and the developer changelog say usage-based billing starts on November 1, 2026, with a free monthly allotment on all Workers plans.

This brief uses Cloudflare's GA blog post, the October 1 changelog entry, and the official limits and pricing docs. No invented benchmarks. No fake customer quotes. Just what delivery teams need for proposals, pilots, and cost notes.

Cloudflare AI Search general availability announcement graphic — Feature image: Cloudflare AI Search GA artwork from the official Cloudflare Blog. Credit: Cloudflare.

What Cloudflare AI Search is, in plain terms

Cloudflare describes AI Search as the search primitive for applications and agents. You create an instance, connect or upload data, and query it with natural language instead of standing up your own chunking, vector database, and hybrid ranker.

For agencies, that sits in a familiar delivery gap. Clients ask for "ChatGPT on our docs," support copilots, or agent tools that can look up internal knowledge. Building that stack from scratch is doable. It also burns weeks before the first useful demo.

Cloudflare AI Search is Cloudflare's managed answer to that gap. It indexes content, runs hybrid retrieval by default on new instances, and can return chunks or feed a generation model for answers. It also exposes MCP and public endpoints that agent builds can call.

What changed at general availability

The GA release is not only a label change. Cloudflare called out several concrete product updates that matter for SOWs.

Native multimodal embedding and retrieval

Older image support relied on object detection plus captions, then text embeddings. That made images findable only through what the caption captured. At GA, AI Search embeds image pixels directly for visual retrieval and still keeps captions for text understanding.

Cloudflare says native multimodal retrieval is available with the Qwen3-VL-Embedding model. The changelog also lists multimodal embedding support for @cf/qwen/qwen3-vl-embedding-2b and google-ai-studio/gemini-embedding-2. Search and chat requests can include images through the REST API and public endpoint.

If an instance uses a text-only embedding model, image queries still work in a fallback path. AI Search converts the query image to text with ToMarkdown and searches on that caption. Native models get the full visual signal.

OCR and larger files

Agencies still receive scanned PDFs from banks, hospitals, manufacturers, and government clients. Text extractors fail on those files. Cloudflare now offers OCR on every account for scanned PDFs.

Plain-text or code files and PDFs with OCR enabled can be up to 10 MiB.
PDFs without OCR and other supported formats remain limited to 4 MiB.
OCR usage is billed as image processing ingestion tokens under the new AI Search pricing.

Supported text-style inputs named in the blog include Markdown, HTML, CSV, JSON, and similar files, plus PDFs. That covers most knowledge-base exports agencies already produce during discovery.

Hybrid search on by default

New AI Search instances use hybrid search by default. Hybrid search combines semantic vector retrieval with full-text matching. You can pick a different index method when you create an instance if a client wants pure vector or keyword behavior.

That default is good for product catalogs, policy manuals, and ticket archives where exact SKUs, error codes, or clause numbers still matter next to fuzzy questions.

Cloudflare AI Search documentation overview graphic — Body image: Cloudflare AI Search docs Open Graph image from developers.cloudflare.com. Credit: Cloudflare.

Cloudflare AI Search pricing agencies can put in a cost sheet

Cloudflare says you pay for three things: content you ingest, data you store, and queries you run. Parsing, chunking, embedding with Workers AI models, keyword indexing, and reranking are included. There are no instance hours or capacity units to size up front.

Billing begins November 1, 2026. Cloudflare will email a reminder the week before. Until then, teams can still enable and use the product.

Monthly free allotment on all Workers plans

5 million ingestion tokens
10 GB-month of storage
1,000 semantic queries
1,000 full-text queries

Image processing tokens share that same 5 million token ingestion pool. The GA blog notes one tweak from earlier preview messaging: free queries are split into 1,000 semantic and 1,000 full-text instead of one shared pool of 2,000.

Published overage rates

Usage	Price
Base ingestion	$0.75 per million tokens
Image processing add-on	+$0.50 per million tokens
Storage	$2.00 per GB-month
Semantic, vector, and hybrid queries	$0.75 per 1,000 queries
Full-text queries	$0.10 per 1,000 queries

Workers AI embedding and reranking calls made by AI Search are included in AI Search pricing. They no longer show up on the Workers AI bill or in AI Gateway logs. Generation, query rewriting, and external providers still bill through your account and gateway.

Ingestion tokens are counted on final chunks after parsing and chunking, using the cl100k_base tokenizer. Chunk overlap is counted in each chunk where it appears. For OCR and images, extracted text also counts as image processing tokens on top of base ingestion.

Cloudflare AI Search GA pricing snapshot for agency planning — Body image: Debugged Pro planning graphic summarizing public Cloudflare AI Search pricing from Oct 1, 2026 docs. Credit: Debugged Pro.

Plan limits that affect client architecture

Pricing is only half the story. Instance and file limits shape whether a pilot or a production knowledge base fits the free Workers plan or needs Workers Paid.

Limit	Workers Free	Workers Paid
AI Search instances per account	100	5,000
Files per instance	100,000	1M, or 500K for hybrid search
Max PDF with OCR / plain-text or code	10 MiB	10 MiB
Max PDF without OCR / other formats	4 MiB	4 MiB
Included semantic queries / month	1,000	1,000
Included full-text queries / month	1,000	1,000
Maximum pages crawled per day	500	Unlimited

Website crawling is bounded by several limits at once. On Workers Free, the 500 pages per day cap is often the binding constraint for large public sites. Paid plans remove that daily crawl ceiling, which matters for continuously updated docs portals.

Custom metadata is capped at five fields per instance. Keep taxonomy simple. Category, locale, product line, audience, and version cover most agency use cases without fighting the limit.

Why this matters for India and USA agency delivery

India delivery centers and USA product teams are hearing the same request: put retrieval behind agents and chat UIs without renting a separate vector database team. Cloudflare AI Search lowers the infra tax when the client already uses Cloudflare for DNS, Workers, R2, or Zero Trust.

Common client asks this product can answer

Internal knowledge search for support and CS teams
Agent tools that query policy, runbooks, and product docs through MCP
Multitenant search where each customer uploads and queries their own files
Website or docs search with embeddable UI snippets
Screenshot and diagram search for product or QA archives

Where agencies still need to do real work

Data cleanup, access rules, and source-of-truth ownership
Prompting, citations, and human review for generated answers
PII redaction before indexing HR, health, or finance content
Cost forecasts once query volume leaves the free 1,000 / 1,000 buckets
Integration with existing auth, SSO, and client network controls

If you already advise clients through software development consulting or broader AI delivery work, treat AI Search as a managed retrieval layer, not a full product. The UI, permissions, evals, and change process still belong in your delivery plan.

Agency pilot checklist for Cloudflare AI Search

Use this sequence on the next India or USA client that asks for RAG without a greenfield platform project.

Confirm the Cloudflare footprint. Check whether the client already has a Workers plan, R2 buckets, or sites on Cloudflare. Source inference can treat an HTTPS URL as a website source or an existing bucket name as an R2 source.
Pick one narrow corpus. Start with a single product docs set, support macro library, or policy pack under a few thousand files. Do not index the whole company drive on day one.
Decide hybrid vs vector vs keyword. Keep hybrid unless the client has a strong reason to change it. Exact IDs and clause numbers usually benefit from the keyword half.
Turn on OCR only where needed. Scanned PDFs need it. Clean digital PDFs and Markdown usually do not. OCR raises image processing token use.
Choose embedding models with the use case. Text-only corpora can stay simple. Screenshot, chart, or product-image search should use a multimodal model Cloudflare lists for AI Search.
Measure answer quality before chat UI polish. Score citation accuracy and "no answer" behavior with a fixed eval set of 30 to 50 real questions from the client.
Model November cost. Estimate tokens after chunk overlap, storage GB, and monthly query mix. Put the free allotment and overage rates in the proposal so finance is not surprised after Nov 1, 2026.
Wire agent access last. Once retrieval quality is stable, connect MCP or the public endpoint to the agent or app your team is building.

For a related security angle on agent delivery, see our recent note on Codex Security Cloud for agency teams. For self-hosted coding agents in regulated environments, see the IBM Bob self-hosted agency guide.

Proposal language you can reuse

Clients do not need a Cloudflare marketing paragraph. They need scope, assumptions, and exits. Here is language that stays honest to public docs.

"We will stand up Cloudflare AI Search against a named document corpus and validate retrieval quality before any generative answer UI goes live."
"OCR will be enabled only for scanned PDFs that fail normal text extraction, because OCR uses image processing ingestion tokens."
"Cost estimates use Cloudflare's published free monthly allotment and overage rates. Billing for AI Search starts November 1, 2026."
"Workers AI embedding and reranking used by AI Search are included in AI Search pricing per Cloudflare's docs. Generation and external models are billed separately."
"If retrieval quality misses the agreed eval threshold, we pause generative answers and continue with search-only or a different corpus split."

Link the client to Cloudflare's own sources in the appendix: the GA blog post, the changelog entry, and the limits and pricing page. That keeps your proposal auditable.

Delivery risks to flag early

Corpus quality beats model choice

Duplicate PDFs, stale Confluence exports, and mixed draft-final folders will poison any managed search product. Budget discovery time for ownership and cleanup. Retrieval quality problems often look like "AI failures" when they are filing cabinet problems.

Query volume can jump after a successful demo

A sales demo with 200 queries looks cheap. A support widget used by every agent all month may not. Put soft alerts around the free 1,000 semantic and 1,000 full-text buckets before November if the client plans a broad rollout.

Multimodal search needs sample queries

If the client cares about screenshot matching or diagram search, collect real query images during discovery. Do not promise multimodal wins from text-only acceptance tests.

Generated answers need a review lane

Cloudflare can return chunks or pass them to a generation model. Agencies still own hallucination risk, citation display, and escalation paths when the index has no good match. Keep a human review step for regulated content.

How Debugged Pro would staff a two-week pilot

A practical India-USA split for a small pilot looks like this.

Day 1-2: confirm Cloudflare account plan, sources, metadata fields, and success metrics with the client product owner.
Day 3-5: ingest the narrow corpus, tune chunking path filters, enable OCR only where scans exist, and build a 40-question eval set.
Day 6-8: run hybrid search evals, compare multimodal vs text-only embeddings if images matter, and freeze retrieval settings.
Day 9-10: attach one consumer (web search UI, support tool, or agent MCP), add citation display, and document cost assumptions for November billing.

That is enough to decide go or no-go before anyone prices a full production knowledge platform. If you want help scoping that pilot, start at our contact page or browse more delivery notes on the Debugged Pro blog.

Key Takeaways

Cloudflare AI Search became generally available on October 1, 2026.
GA adds native image embeddings, OCR for scanned PDFs, and 10 MiB limits for OCR-enabled PDFs and plain-text or code files.
Usage-based billing starts November 1, 2026, with reminder email the week before.
Free monthly allotment includes 5M ingestion tokens, 10 GB-month storage, 1,000 semantic queries, and 1,000 full-text queries.
New instances default to hybrid search; Workers AI embedding and reranking used by AI Search are included in AI Search pricing.
India and USA agencies can use it to shorten RAG and agent-search pilots when clients already sit on Cloudflare, but corpus cleanup, evals, and answer review remain agency work.

About the author: Nitin Kumar is Lead Project Coordinator at Debugged Pro. He helps India and USA software agencies turn vendor releases into delivery plans clients can approve, staff, and ship.