A server that is not ready does not always take its time about saying so.
Ollama answers 5xx immediately while a model is still being loaded, and a
gateway with no upstream yet does the same, so the failure arrives at
once rather than as a timeout.
Both mean "not right now" and are worth one more attempt. 4xx keeps
being final -- a wrong API key or an unknown model name is not going to
change between two requests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An idle LLM server unloads the model and loads it again on the next
request. Nothing of the response is sent while that load runs, so a
request that arrives on a cold server can exceed read_timeout and fail --
reliably making the first search after an idle period the one that does
not get a summary. That same request is what starts the load, so asking
again succeeds.
Only a server that failed to answer is asked again. A server that did
answer with an error status is not: a wrong API key or an unknown model
name does not become right on a second attempt.
The status and body of such an error response are now logged. Until now
every upstream failure looked identical from the outside -- HTTP 502 with
no indication of whether the key, the model name or the network was at
fault.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three changes to the ai_summary plugin:
- The *AI Summary* preferences tab is only rendered when the plugin is
activated in settings.yml. An instance that does not offer AI
summaries no longer shows an AI tab at all. The gate is the
administrator setting, not the user opt-out, because the per user
on/off switch lives inside that tab -- hiding it on opt-out would
leave no way to opt back in.
- Users can configure an API key for their own LLM server
(ai_summary_api_key). The administrator key is still only sent to
base_url and the user key only to a server the user configured, so
neither key can be captured through the other. The setting is marked
secret: credentials are excluded from the preferences URL, which users
copy around to transfer or share their preferences.
- Grounding summaries on the search results is now the default; the
extra cost of the longer prompt is moderate.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
httpx derives an "Authorization: Basic" header from the userinfo of a
URL, so a user could make SearXNG send a header of their choosing to a
host of their choosing (e.g. to probe an internal service behind basic
auth). A user preference carrying credentials is now ignored and the
administrator default is used instead; credentials in the configured
base_url are untouched.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Servers that require authentication (e.g. vLLM or llama.cpp started with
--api-key, or an LLM server behind an authenticating reverse proxy) can
now be configured with an ai_summary.api_key, sent as "Authorization:
Bearer".
The key is administrator configuration only: there is no preference for
it, and it is only sent to the configured base_url. Users can point the
ai_summary_server preference at a server of their own, and such a server
must not be handed the instance API key -- otherwise every user of the
instance could capture it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Talk to the LLM server via GET /v1/models and POST /v1/chat/completions
(SSE) instead of Ollama's native API. Any OpenAI compatible server now
works (Ollama, vLLM, llama.cpp, LM Studio, Hugging Face TGI, ...);
Ollama serves this API natively, existing setups keep working unchanged.
The Ollama specific keep_alive option is dropped, the ai_summary.grounding
setting is added as instance wide default of the grounding preference.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add an optional, disabled-by-default plugin that shows an AI generated
summary at the top of the result page, generated by a (local) Ollama
server:
- async: the result page is never delayed; a client plugin streams the
answer (NDJSON over a new /ai_summary endpoint) into a placeholder
answer with a typing indicator, collapsed behind a More button, with
an inline follow-up chat
- trigger: first page of general searches only, skipped when an infobox
or instant answer already answers the query
- grounding (per-user preference): send the top result snippets as
context, the model answers from them instead of its own knowledge
- configuration: new AI Summary preferences tab (server URL, model,
grounding) with instance defaults in a new ai_summary: settings
section; all three preferences can be locked for public instances
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
jasonwitty