Web Search, Web Pages and Scanned PDFs: How Aikeedo 5.2 Reads the World

The smartest model gives a poor answer when it cannot see the page it was asked about. A search that returns noise, a site that only renders with JavaScript, a contract that was scanned instead of exported, a spreadsheet the app would not accept — each one ends with the AI guessing or apologizing.

v5.2 rebuilds the three ways Aikeedo brings outside information into a conversation: searching the web, reading web pages and reading files. For each, you now choose the service that does the work, and you can keep your own server as the first choice with a paid service as the fallback.

Everything below lives under Settings → Features → Chat.

Pick Your Search Provider

Web search is now a choice of provider rather than a fixed integration:

• Jina — search through Jina's API, with the same key as its reader
• Serper — Google results through Serper
• Perplexity — Perplexity's Search API, in Standard or Fast mode

You also decide how much each search adds to the conversation. Results per search runs from 1 to 20, with 5 as the default, and Search result size caps how many characters a search can add. More results give the model more to go on; fewer keep requests cheap.

Results now reach the model as a short numbered list of titles, links and snippets instead of the raw data a provider returns. For a typical Serper search that is about 64% less input than the raw results Aikeedo used to pass along, and your users notice it in their credit balance. Perplexity results also include page excerpts, so the model often has what it needs without opening each link.

The Perplexity integration covers search. If you want your users to chat with Perplexity's models, add Perplexity as a custom LLM server.

Reading Web Pages

When a user pastes a link, or a search turns up a page worth reading, Aikeedo opens it. The Web page reader setting chooses how:

• Built-in — reads the page on your own server, at no extra cost. It now keeps tables as rows and keeps paragraph breaks, so the model sees structure instead of a wall of text.
• Jina — also renders pages that need JavaScript.
• Cloudflare — Cloudflare Browser Rendering loads the page in a real browser.

With Jina or Cloudflare selected, Use only as fallback is on by default: your server reads every page first, and the remote service steps in only for pages it cannot read. Turn it off to send every page to the remote service first. Web page size caps how much of a long page the model gets.

Every address a user supplies is checked before anything fetches it, so links to your server's own network are refused — the same protection we added for connectors in v5.1. The same readers are used for URLs added to an assistant's knowledge.

Reading Files

The File reader setting works the same way:

• Built-in — reads files on your server.
• Jina — adds OCR for scanned PDFs.
• Mistral OCR — reads scanned PDFs, tables and handwriting, and returns them as Markdown that keeps the layout. It is billed per page.

With Use only as fallback on, your server reads each file first and the remote reader handles only the ones it cannot, such as a PDF that is really a stack of scanned images. Files sent to Jina or Mistral are processed by those services, which is worth mentioning in your terms.

If no reader can make sense of a file, chat still attaches it and tells the model it could not be read, while a knowledge source is rejected with an explanation. Nothing fails silently.

More formats

Chat attachments and knowledge sources now accept XLSX and ODS spreadsheets (one table per sheet), PPTX presentations (text slide by slide), RTF and Markdown, alongside PDF, Word, ODT, CSV, JSON, XML, YAML and plain text. The older .xls format is not supported; save it as XLSX first.

The Model Sees the Whole File

Until now, an attached file was cut into pieces and only the pieces that matched the question reached the model. That works for finding a fact. It does not work for "summarize this" or "compare these two contracts".

In v5.2, small files go to the model whole. The Read in full up to setting sets the limit — 8,000 tokens by default, about 32,000 characters — and 0 switches back to searching in pieces every time.

Larger files are listed for the model, which can then read them page by page with a new read_file tool, or search them as before. A long report is no longer reduced to a few matching paragraphs.

Safer Uploads

Every file is checked before it is stored. Documents must be one of the accepted types, and both documents and images are checked against their actual content, not just the file name. A file that fails is rejected with the reason.

Max file size sets the largest file users can attach in chat. Leave it empty to allow anything up to your server's PHP limits.

Credits

Each new service has its own credit rate, so you can price it like everything else on your plans:

• Jina Search and Jina Reader
• Perplexity Search, with separate rates for Standard and Fast
• Cloudflare Browser Rendering, per page
• Mistral OCR, per page

The built-in reader costs nothing beyond the model's tokens, which is one more reason the fallback setting is on by default.

Upgrading: Turn Web Search Back On

The search tool was rebuilt for v5.2, and its old settings do not carry over. After updating, web search is off until you set it up again:

1. Under Settings → Features → Chat, choose a Search provider. The default is none.
2. Add that provider's API key on its settings page.
3. Turn on web search in each plan that should include it, and update the plan snapshots so existing subscribers get it.

Readers need nothing: Built-in is the default, and you can add Jina, Cloudflare or Mistral whenever you like.

Getting Started

The web search and reading documentation and the file reading documentation walk through every setting, with setup guides for Jina, Perplexity, Cloudflare Browser Rendering and Mistral OCR. The feature page has the overview, and the v5.2 release post covers everything else that is new.

Try the live demo to see it in action, or head to our pricing page if you are ready to get started.