Enabling Web Search and Image Generation

This page shows how to let the AI search the internet (web search, internet search, sources as links) and generate images (image generation, DALL·E-style models) — both on the third card of the settings page General.

What this does

• Web search lets the AI look things up on the web when a question needs current information; it lists its sources as links. You choose the search service (Tavily, Brave Search, Serper or a self-hosted SearXNG) and provide its API key.
• Image generation lets the AI create an image on request, using an image model of one of your connections.

Both are optional, apply to Jira and Confluence alike, and are independent of the AI access to Jira & Confluence policy.

Who can do this

Product administrators (Jira or Confluence administration) on the settings page AnyLLM Chat — Settings — see Opening the Settings and Initial Setup. Approving the search endpoint follows the rules of Managing Endpoint Approvals (meant for Jira administrators).

Before you start

• Web search: an account and API key with the search service of your choice — or a SearXNG instance of your own with the JSON output format enabled.
• Image generation: an OpenAI-compatible or Azure OpenAI connection whose service offers an image model.

Enabling web search

Purpose — allow the AI to search the web and cite sources.

Steps

1. Open Settings → General and scroll to the card 🌐 Web search.
2. Tick Enable — the AI may search the web when needed (sources as links).
3. Choose the Search service. Choosing a service fills Endpoint (https, base URL without path) with that service's preset address and clears any key you have typed.
4. Check the endpoint. For SearXNG (self-hosted) enter the address of your own instance, e.g. https://search.my-company.com.
5. Enter the API key of the search service (at least 8 characters). The field is hidden for SearXNG. If a key is already stored, the label reads (set — leave empty to keep).
6. Click Save. If the endpoint address is new, the Atlassian confirmation dialog opens — confirm it.
7. Open the settings in the other product as well and approve the same address there (Endpoint approvals → Approval status per product → Approve now), because the administrator's consent counts per product.

Expected result — the status reads Saved. From now on the AI may call the search when it sees fit; a search shows up in the chat as a tool step row Web search: …, and the answer names sources as links.

Available services

Notes

• Like an LLM endpoint, the search endpoint needs an approval in each product — the preset addresses included.
• Switching the search service or changing the endpoint address discards the stored key, so that it is never sent to a different service. Enter the key again afterwards. Only if service and address stay unchanged does "leave empty to keep" apply.
• Without an endpoint the search stays off even with the checkbox ticked. If the key is missing for a service that requires one, the search tool is deliberately not offered to the AI.
• With SearXNG, the JSON output format must be enabled in your instance; the app calls it without a key.
• At most 5 results feed into an answer; a search is aborted after 30 seconds. The query is cut at 400 characters and each result excerpt at 1,200 characters.
• Brave Search, Serper and SearXNG were built from the providers' documentation and have not been live-tested by the vendor — after setting one of them up, ask a question that triggers a search to confirm that address, approval and key fit together.
• Search errors are returned to the AI, not shown to you directly; the AI may paraphrase them in its answer. The tool step row Web search: … shows what was attempted.

Common problems

Enabling image generation

Purpose — allow the AI to generate images on request.

Steps

1. Under Connections, add the image model (e.g. dall-e-3, flux) to the Models (comma-separated) list of an OpenAI-compatible or Azure OpenAI connection and click Save connections (Setting Up LLM Connections).
2. Open Settings → General and scroll to the card 🎨 Image generation.
3. Tick Enable — the AI may generate images (/images/generations).
4. Under Image model (via one of the connections), replace — choose — with the image model; the list groups the models of all saved connections by connection.
5. Click Save.

Expected result — the status reads Saved. When a user asks for an image, the chat shows the tool step row Generating image: … and the AI embeds the image in its answer via the image address (URL) returned by the service.

Notes

• Image generation works only with a model on an OpenAI-compatible (incl. self-hosted) or Azure OpenAI connection. The selector also lists models of Anthropic connections, but an image model chosen there is silently not offered to the AI — users get no error, just no images.
• The image prompt is limited to 2,000 characters, and the service must return an image URL (services that only return embedded image data are not supported).
• Image links from most providers are valid for a short time only; users who want to keep an image should download it promptly.
• Errors are returned to the AI, not shown directly; the AI may paraphrase them.

Related topics

• Administration & Configuration
• General Settings: Language, Default Model, AI Access, Rate Limit, Site Knowledge
• Setting Up LLM Connections
• Managing Endpoint Approvals
• Letting the AI Search, Read and Create in Jira and Confluence — how users experience web search and images
• Troubleshooting
• FAQ