[feat] plugin: AI summary of search results via local Ollama server

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>

client/simple/src/js/plugin/AiSummary.ts

@@ -0,0 +1,218 @@
+// SPDX-License-Identifier: AGPL-3.0-or-later
+
+import { Plugin } from "../Plugin.ts";
+import { settings } from "../toolkit.ts";
+import { assertElement } from "../util/assertElement.ts";
+
+type Message = { role: "user" | "assistant"; content: string };
+type ContextItem = { title: string; url: string; snippet: string };
+type StreamLine = { delta?: string; done?: boolean; model?: string; error?: string };
+
+// keep in sync with the server side defaults (searx/ai_summary.py)
+const MAX_CONTEXT_ITEMS = 5;
+const MAX_HISTORY_MESSAGES = 12;
+const MAX_TITLE_LENGTH = 300;
+const MAX_SNIPPET_LENGTH = 1000;
+
+/**
+ * Fills the AI summary placeholder (rendered by the ai_summary plugin of the
+ * server) by streaming an answer from the /ai_summary endpoint, with an
+ * expand button and a follow-up question chat.
+ */
+export default class AiSummary extends Plugin {
+  private readonly messages: Message[] = [];
+  private context: ContextItem[] = [];
+  private controller: AbortController | undefined;
+
+  public constructor() {
+    super("ai_summary");
+  }
+
+  protected async run(): Promise<void> {
+    const box = document.getElementById("ai_summary");
+    if (!box) return;
+
+    try {
+      const { query } = box.dataset;
+      if (!query) return;
+
+      if (box.dataset.grounding === "1") {
+        this.context = AiSummary.collectContext();
+      }
+
+      this.wireControls(box);
+
+      this.messages.push({ role: "user", content: query });
+      await this.exchange(box);
+    } catch (error) {
+      // never fail silently, always leave a message in the summary box
+      AiSummary.showError(box, error);
+    }
+  }
+
+  protected async post(): Promise<void> {
+    // noop
+  }
+
+  /**
+   * Scrape the top search results from the DOM as grounding context.
+   */
+  private static collectContext(): ContextItem[] {
+    const items: ContextItem[] = [];
+    for (const article of document.querySelectorAll<HTMLElement>("#urls article.result")) {
+      if (items.length >= MAX_CONTEXT_ITEMS) break;
+
+      const link = article.querySelector<HTMLAnchorElement>("h3 a");
+      if (!link) continue;
+
+      items.push({
+        title: (link.textContent ?? "").trim().slice(0, MAX_TITLE_LENGTH),
+        url: link.href,
+        snippet: (article.querySelector(".content")?.textContent ?? "").trim().slice(0, MAX_SNIPPET_LENGTH)
+      });
+    }
+    return items;
+  }
+
+  private wireControls(box: HTMLElement): void {
+    const body = box.querySelector<HTMLElement>(".ai-summary-body");
+    const moreButton = box.querySelector<HTMLElement>(".ai-summary-more");
+    const followupForm = box.querySelector<HTMLFormElement>(".ai-summary-followup");
+    assertElement(body);
+    assertElement(moreButton);
+    assertElement(followupForm);
+
+    moreButton.addEventListener("click", () => {
+      const collapsed = body.classList.toggle("collapsed");
+      moreButton.textContent = collapsed
+        ? (moreButton.dataset.btnTextCollapsed ?? "")
+        : (moreButton.dataset.btnTextNotCollapsed ?? "");
+      followupForm.classList.toggle("invisible", collapsed);
+    });
+
+    followupForm.addEventListener("submit", (event: Event) => {
+      event.preventDefault();
+
+      const input = followupForm.querySelector<HTMLInputElement>("input");
+      assertElement(input);
+
+      const question = input.value.trim();
+      if (!question || this.controller) return;
+
+      input.value = "";
+      this.messages.push({ role: "user", content: question });
+      // the server rejects too long histories, drop the oldest messages
+      this.messages.splice(0, this.messages.length - (MAX_HISTORY_MESSAGES - 1));
+
+      const questionElement = Object.assign(document.createElement("p"), {
+        textContent: question,
+        className: "ai-summary-question"
+      });
+      box.querySelector(".ai-summary-answers")?.append(questionElement);
+
+      void this.exchange(box);
+    });
+  }
+
+  /**
+   * Send the message history to /ai_summary and stream the answer into a new
+   * block in the answer area.
+   */
+  private async exchange(box: HTMLElement): Promise<void> {
+    const answers = box.querySelector<HTMLElement>(".ai-summary-answers");
+    assertElement(answers);
+
+    const block = Object.assign(document.createElement("p"), {
+      className: "ai-summary-content typing"
+    });
+    answers.append(block);
+
+    const controller = new AbortController();
+    this.controller = controller;
+
+    let text = "";
+    const handleLine = (line: string): void => {
+      if (!line.trim()) return;
+
+      const data = JSON.parse(line) as StreamLine;
+      if (data.error) {
+        throw new Error(data.error);
+      }
+      if (data.delta) {
+        text += data.delta;
+        block.textContent = text;
+        this.updateMoreButton(box);
+      }
+      if (data.done && data.model) {
+        const modelElement = box.querySelector<HTMLElement>(".ai-summary-model");
+        if (modelElement) modelElement.textContent = data.model;
+      }
+    };
+
+    try {
+      const res = await fetch("./ai_summary", {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ messages: this.messages, context: this.context }),
+        signal: controller.signal
+      });
+      if (!res.ok) {
+        throw new Error(res.statusText);
+      }
+
+      if (res.body) {
+        const reader = res.body.getReader();
+        const decoder = new TextDecoder();
+        let buffer = "";
+
+        for (;;) {
+          // biome-ignore lint/performance/noAwaitInLoops: chunks of a stream are read sequentially
+          const { done, value } = await reader.read();
+          if (done) break;
+
+          buffer += decoder.decode(value, { stream: true });
+          const lines = buffer.split("\n");
+          buffer = lines.pop() ?? "";
+          for (const line of lines) handleLine(line);
+        }
+        handleLine(buffer);
+      } else {
+        // some webviews (e.g. Brave iOS) wrap fetch and don't expose a
+        // streaming body: no typing effect, render the answer in one go
+        const body = await res.text();
+        for (const line of body.split("\n")) handleLine(line);
+      }
+      this.messages.push({ role: "assistant", content: text });
+    } catch (error) {
+      AiSummary.showError(box, error);
+    } finally {
+      block.classList.remove("typing");
+      this.controller = undefined;
+      this.updateMoreButton(box);
+    }
+  }
+
+  private static showError(box: HTMLElement, error: unknown): void {
+    console.error("Error loading AI summary:", error);
+
+    const errorElement = Object.assign(document.createElement("div"), {
+      textContent: settings.translations?.error_loading_ai_summary ?? "Error loading the AI summary",
+      className: "dialog-error"
+    });
+    errorElement.setAttribute("role", "alert");
+    (box.querySelector<HTMLElement>(".ai-summary-answers") ?? box).append(errorElement);
+  }
+
+  /**
+   * Show the expand button as soon as the (collapsed) body overflows.
+   */
+  private updateMoreButton(box: HTMLElement): void {
+    const body = box.querySelector<HTMLElement>(".ai-summary-body");
+    const moreButton = box.querySelector<HTMLElement>(".ai-summary-more");
+    if (!(body && moreButton)) return;
+
+    if (!body.classList.contains("collapsed") || body.scrollHeight > body.clientHeight + 4) {
+      moreButton.classList.remove("invisible");
+    }
+  }
+}

client/simple/src/js/router.ts

@@ -34,6 +34,13 @@ ready(() => {
       where: [Endpoints.results]
     });
   }
+
+  if (settings.plugins?.includes("ai_summary")) {
+    load(() => import("./plugin/AiSummary.ts").then(({ default: Plugin }) => new Plugin()), {
+      on: "endpoint",
+      where: [Endpoints.results]
+    });
+  }
 });
 
 ready(

client/simple/src/less/ai_summary.less

@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: AGPL-3.0-or-later
+
+@keyframes ai-summary-caret {
+  50% {
+    opacity: 0;
+  }
+}
+
+.ai-summary {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+
+  .ai-summary-header {
+    display: flex;
+    align-items: baseline;
+    gap: 0.5rem;
+
+    .ai-summary-title {
+      font-size: 0.9em;
+      font-weight: bold;
+      text-transform: uppercase;
+      letter-spacing: 0.05em;
+    }
+
+    .ai-summary-model {
+      font-size: 0.8em;
+      opacity: 0.7;
+    }
+  }
+
+  .ai-summary-body.collapsed {
+    max-height: 8em;
+    overflow: hidden;
+    mask-image: linear-gradient(to bottom, #000 60%, transparent 100%);
+  }
+
+  .ai-summary-answers {
+    display: flex;
+    flex-direction: column;
+    gap: 0.5rem;
+  }
+
+  .ai-summary-content {
+    margin: 0;
+    white-space: pre-line;
+    overflow-wrap: anywhere;
+
+    &.typing::after {
+      content: "▍";
+      animation: ai-summary-caret 1s step-end infinite;
+    }
+  }
+
+  .ai-summary-question {
+    margin: 0;
+    font-weight: bold;
+  }
+
+  .ai-summary-more {
+    align-self: center;
+    border: none;
+    .show-content-button();
+  }
+
+  .ai-summary-followup {
+    display: flex;
+    gap: 0.5rem;
+
+    input {
+      flex: 1;
+      padding: 5px 10px;
+      border: 1px solid var(--color-sidebar-border);
+      background: var(--color-answer-background);
+      color: var(--color-answer-font);
+      .rounded-corners-tiny;
+    }
+
+    button {
+      border: none;
+      .show-content-button();
+    }
+  }
+
+  .ai-summary-disclaimer {
+    margin: 0;
+    font-size: 0.8em;
+    opacity: 0.7;
+  }
+}

client/simple/src/less/style.less

@@ -13,6 +13,7 @@
 @import "stats.less";
 @import "result_templates.less";
 @import "weather.less";
+@import "ai_summary.less";
 
 // for index.html template
 @import "index.less";

docs/admin/settings/index.rst

@@ -25,3 +25,4 @@ Settings
    settings_outgoing
    settings_categories_as_tabs
    settings_plugins
+   settings_ai_summary

docs/admin/settings/settings_ai_summary.rst

@@ -0,0 +1,36 @@
+.. _settings ai_summary:
+
+===============
+``ai_summary:``
+===============
+
+Default configuration of the :ref:`AI summary plugin <ai_summary plugin>`.
+Users configure the Ollama server URL and the model in the *AI Summary* tab of
+their preferences; the values below only act as instance wide defaults.
+
+.. code:: yaml
+
+   ai_summary:
+     base_url: "http://127.0.0.1:11434"
+     model: "llama3.2:3b"
+
+.. attention::
+
+   A user configurable server URL allows any user of the instance to make the
+   SearXNG server send requests to a URL of their choice (`SSRF`_), and each
+   summary is real LLM work.  This plugin is intended for private instances --
+   on a public instance lock the related preferences (:ref:`settings
+   preferences`):
+
+   .. code:: yaml
+
+      preferences:
+        lock:
+          - ai_summary_server
+          - ai_summary_model
+          - ai_summary_grounding
+
+.. _SSRF: https://owasp.org/www-community/attacks/Server_Side_Request_Forgery
+
+.. autoclass:: searx.ai_summary.SettingsAISummary
+   :members:

docs/dev/plugins/ai_summary.rst

@@ -0,0 +1,8 @@
+.. _ai_summary plugin:
+
+==========
+AI summary
+==========
+
+.. automodule:: searx.plugins.ai_summary
+   :members:

docs/dev/plugins/builtins.rst

@@ -7,6 +7,7 @@ Built-in Plugins
 .. toctree::
    :maxdepth: 1
 
+   ai_summary
    calculator
    hash_plugin
    hostnames

searx/_settings.py

@@ -33,6 +33,9 @@ class SettingsPref(msgspec.Struct, kw_only=True, forbid_unknown_fields=True):
             "theme",
             "results_on_new_tab",
             "doi_resolver",
+            "ai_summary_server",
+            "ai_summary_model",
+            "ai_summary_grounding",
             "simple_style",
             "center_alignment",
             "query_in_title",

searx/ai_summary.py

@@ -0,0 +1,121 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+"""Implementations needed for the AI summary plugin
+(:py:obj:`searx.plugins.ai_summary`)."""
+# pylint: disable=too-few-public-methods
+
+# Struct fields aren't discovered in Python 3.14
+# - https://github.com/searxng/searxng/issues/5284
+from __future__ import annotations
+
+__all__ = ["SettingsAISummary", "MODELS", "model_choices", "build_ollama_messages"]
+
+import msgspec
+
+DEFAULT_SYSTEM_PROMPT = (
+    "You are a search assistant.  Answer the user's search query concisely in"
+    " a few short paragraphs of plain text.  If you are unsure or don't know"
+    " the answer, say so."
+)
+
+DEFAULT_SYSTEM_PROMPT_GROUNDED = (
+    "You are a search assistant.  Answer the user's search query concisely in"
+    " a few short paragraphs of plain text, using the following search results"
+    " as context when they are relevant.  If you are unsure or don't know the"
+    " answer, say so.\n\nSearch results:\n\n{context}"
+)
+
+MODELS: list[str] = []
+"""List of model names a user can select from.  Populated once at
+application setup by :py:obj:`searx.plugins.ai_summary.SXNGPlugin.init`."""
+
+
+class SettingsAISummary(msgspec.Struct, kw_only=True, forbid_unknown_fields=True):
+    """Options for configuring the AI summary plugin.
+
+    .. code:: yaml
+
+       ai_summary:
+         base_url: "http://127.0.0.1:11434"
+         model: "llama3.2:3b"
+         models:
+           - "llama3.2:3b"
+           - "gemma3:4b"
+    """
+
+    base_url: str = ""
+    """Default base URL of the Ollama server (e.g. ``http://127.0.0.1:11434``).
+    Users can set their own server URL in the preferences
+    (``ai_summary_server``) unless that preference is locked."""
+
+    model: str = ""
+    """Name of the default model (e.g. ``llama3.2:3b``).  If empty, the first
+    entry of :py:obj:`SettingsAISummary.models` is used.  Users can set their
+    own model in the preferences (``ai_summary_model``) unless that preference
+    is locked."""
+
+    models: list[str] = []
+    """List of model names suggested to the user in the preferences.  If empty
+    and :py:obj:`SettingsAISummary.base_url` is set, the list is requested
+    once at application setup from the Ollama server (``GET /api/tags``)."""
+
+    grounding: bool = False
+    """Default of the ``ai_summary_grounding`` user preference: ground the
+    summary on the search results.  Users can still opt in/out in the
+    preferences unless that preference is locked."""
+
+    connect_timeout: float = 5.0
+    """Timeout (seconds) to establish a TCP connection to the Ollama server."""
+
+    read_timeout: float = 30.0
+    """Maximum gap (seconds) between two chunks of the token stream."""
+
+    stream_timeout: float = 120.0
+    """Wall clock limit (seconds) for one completion."""
+
+    keep_alive: str = "5m"
+    """How long the model stays loaded in memory after the request (passed
+    through to Ollama's ``keep_alive`` option)."""
+
+    max_context_items: int = 5
+    """Maximum number of search results accepted as grounding context."""
+
+    max_history_messages: int = 12
+    """Maximum number of messages (follow-up chat history) per request."""
+
+    max_message_length: int = 4000
+    """Maximum length (characters) of a single message or context snippet."""
+
+    system_prompt: str = DEFAULT_SYSTEM_PROMPT
+    """System prompt used when the *grounding* preference is off."""
+
+    system_prompt_grounded: str = DEFAULT_SYSTEM_PROMPT_GROUNDED
+    """System prompt used when the *grounding* preference is on.  The
+    placeholder ``{context}`` is replaced by an enumeration of the search
+    results sent along with the query."""
+
+
+def model_choices() -> list[str]:
+    """Model names a user can select from in the preferences."""
+    return list(MODELS)
+
+
+def build_ollama_messages(
+    cfg: SettingsAISummary,
+    messages: list[dict[str, str]],
+    context: list[dict[str, str]] | None = None,
+) -> list[dict[str, str]]:
+    """Build the message list for Ollama's ``/api/chat`` from the (already
+    validated) request ``messages``, prepending a system prompt.  When
+    ``context`` items are given, the grounded system prompt is used and the
+    context items are serialized into its ``{context}`` placeholder."""
+
+    if context:
+        ctx_lines = [
+            f"[{no}] {item.get('title', '')} — {item.get('snippet', '')} ({item.get('url', '')})"
+            for no, item in enumerate(context[: cfg.max_context_items], start=1)
+        ]
+        system_prompt = cfg.system_prompt_grounded.replace("{context}", "\n".join(ctx_lines))
+    else:
+        system_prompt = cfg.system_prompt
+
+    return [{"role": "system", "content": system_prompt}, *messages]

searx/plugins/_core.py

@@ -42,7 +42,7 @@ class PluginInfo:
     description: str
     """Short description of the *answerer*."""
 
-    preference_section: t.Literal["general", "ui", "privacy", "query"] | None = "general"
+    preference_section: t.Literal["general", "ui", "privacy", "query", "ai"] | None = "general"
     """Section (tab/group) in the preferences where this plugin is shown to the
     user.
 

searx/plugins/ai_summary.py

@@ -0,0 +1,302 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+"""Plugin that displays an AI generated summary of the search query at the top
+of the result page.  The summary is generated by a (local) `Ollama`_ server.
+
+The Ollama server URL and the model are configured by the user in the *AI
+Summary* tab of the preferences (``ai_summary_server``, ``ai_summary_model``);
+the administrator can configure instance wide defaults in the ``ai_summary:``
+section and lock the preferences via :ref:`settings preferences`.
+
+.. attention::
+
+   A user configurable server URL allows any user of the instance to make the
+   SearXNG server send requests to a URL of their choice (`SSRF`_), and each
+   summary is real LLM work.  This plugin is intended for private instances --
+   on a public instance, lock the ``ai_summary_server``, ``ai_summary_model``
+   and ``ai_summary_grounding`` preferences and configure the ``ai_summary:``
+   section instead.
+
+The result page is never delayed by this plugin: it only places an empty
+placeholder (:py:obj:`searx.result_types.AiSummary`) in the answer area, which
+is filled asynchronously by the client (``client/simple/src/js/plugin/
+AiSummary.ts``) from the ``/ai_summary`` endpoint (registered in
+:py:obj:`searx.webapp`).  The endpoint streams the tokens from Ollama's
+``/api/chat`` to the client as `NDJSON`_.
+
+A summary is only generated on the first page of a *general* search and only
+if no engine has contributed an infobox (e.g. wikipedia / wikidata) or an
+instant answer (e.g. ddg definitions) -- in these cases the query is most
+likely a lookup of a well known term that is already answered.
+
+The requests to the Ollama server are sent directly (not via
+:py:obj:`searx.network`), an outgoing proxy configuration is deliberately not
+applied to reach an Ollama server in the local network.
+
+Configuration of the defaults (:py:obj:`searx.ai_summary.SettingsAISummary`):
+
+.. code:: yaml
+
+   ai_summary:
+     base_url: "http://127.0.0.1:11434"
+     model: "llama3.2:3b"
+
+.. code:: yaml
+
+   plugins:
+     searx.plugins.ai_summary.SXNGPlugin:
+       active: false
+
+.. _Ollama: https://ollama.com/
+.. _NDJSON: https://github.com/ndjson/ndjson-spec
+.. _SSRF: https://owasp.org/www-community/attacks/Server_Side_Request_Forgery
+"""
+
+import typing as t
+
+import json
+import logging
+import re
+import time
+from urllib.parse import urlparse
+
+import flask
+import httpx
+from flask_babel import gettext
+
+from searx import get_setting
+from searx.ai_summary import SettingsAISummary, build_ollama_messages
+from searx.extended_types import sxng_request
+from searx.result_types import EngineResults
+import searx.ai_summary
+
+from . import Plugin, PluginInfo
+
+if t.TYPE_CHECKING:
+    from searx.search import SearchWithPlugins
+    from searx.extended_types import SXNG_Request
+    from . import PluginCfg
+
+VALID_ROLES = ("user", "assistant")
+MODEL_NAME_REGEXP = re.compile(r"[A-Za-z0-9._:/-]{1,128}")
+
+
+def _get_client(base_url: str, cfg: SettingsAISummary) -> httpx.Client:
+    """HTTP client for one request to the Ollama server at ``base_url``."""
+    return httpx.Client(
+        base_url=base_url,
+        timeout=httpx.Timeout(connect=cfg.connect_timeout, read=cfg.read_timeout, write=10.0, pool=10.0),
+    )
+
+
+def _valid_server(url: str) -> bool:
+    try:
+        parsed = urlparse(url)
+    except ValueError:
+        return False
+    return parsed.scheme in ("http", "https") and bool(parsed.netloc) and len(url) <= 256
+
+
+def _user_server(request: "SXNG_Request", cfg: SettingsAISummary) -> str:
+    """The Ollama server URL for this request: the user's ``ai_summary_server``
+    preference, or the administrator's default."""
+    return str(request.preferences.get_value("ai_summary_server") or "").strip() or cfg.base_url
+
+
+class SXNGPlugin(Plugin):
+    """Plugin that adds the AI summary placeholder to the result page, the
+    ``/ai_summary`` endpoint itself is registered in :py:obj:`searx.webapp`."""
+
+    id = "ai_summary"
+
+    def __init__(self, plg_cfg: "PluginCfg"):
+        super().__init__(plg_cfg)
+
+        self.info = PluginInfo(
+            id=self.id,
+            name=gettext("AI summary"),
+            description=gettext(
+                "Show an AI generated summary of the search query on top of the"
+                " result page (uses an Ollama server, see the settings below)."
+            ),
+            preference_section="ai",
+        )
+
+    def init(self, app: "flask.Flask") -> bool:
+        cfg: SettingsAISummary = get_setting("ai_summary")
+
+        if cfg.base_url:
+            searx.ai_summary.MODELS = list(cfg.models) or self._probe_models(cfg)
+            if not cfg.model and searx.ai_summary.MODELS:
+                cfg.model = searx.ai_summary.MODELS[0]
+            if cfg.model and cfg.model not in searx.ai_summary.MODELS:
+                searx.ai_summary.MODELS.insert(0, cfg.model)
+
+        return True
+
+    def _probe_models(self, cfg: SettingsAISummary) -> list[str]:
+        """Request the list of models from the Ollama server (``GET
+        /api/tags``).  The server might not be up when SearXNG starts, a
+        failing probe only leaves the model suggestion list empty."""
+        try:
+            with _get_client(cfg.base_url, cfg) as client:
+                resp = client.get("/api/tags")
+                resp.raise_for_status()
+                models = [model["name"] for model in resp.json().get("models", [])]
+        except (httpx.HTTPError, ValueError, KeyError) as exc:
+            self.log.warning("can't request model list from %s: %s", cfg.base_url, exc)
+            models = []
+        return models or ([cfg.model] if cfg.model else [])
+
+    def post_search(self, request: "SXNG_Request", search: "SearchWithPlugins") -> EngineResults | None:
+        results = EngineResults()
+        sq = search.search_query
+        cfg: SettingsAISummary = get_setting("ai_summary")
+
+        skip = (
+            sq.pageno > 1
+            # post_search is also called for the json, csv and rss formats,
+            # the placeholder is only useful on the HTML result page
+            or request.form.get("format", "html") != "html"
+            or "general" not in sq.categories
+            # an infobox (e.g. wikipedia / wikidata) or an instant answer
+            # (e.g. ddg definitions) most likely already answers the query
+            or bool(search.result_container.infoboxes)
+            or bool(search.result_container.answers)
+            or not sq.query.strip()
+            # without an Ollama server (user preference or admin default)
+            # there is nothing to show
+            or not _user_server(request, cfg)
+        )
+        if skip:
+            return None
+
+        grounding = bool(request.preferences.get_value("ai_summary_grounding"))
+        results.add(results.types.AiSummary(query=sq.query, grounding=grounding))
+        return results
+
+
+def _bad_request(msg: str) -> flask.Response:
+    return flask.Response(json.dumps({"error": msg}), status=400, mimetype="application/json")
+
+
+def _validate_messages(messages: t.Any, cfg: SettingsAISummary) -> list[dict[str, str]]:
+    if not isinstance(messages, list) or not messages or len(messages) > cfg.max_history_messages:
+        raise ValueError("invalid messages")
+    for msg in messages:
+        if not isinstance(msg, dict) or msg.keys() != {"role", "content"}:
+            raise ValueError("invalid message")
+        if msg["role"] not in VALID_ROLES or not isinstance(msg["content"], str):
+            raise ValueError("invalid message")
+        if not msg["content"].strip() or len(msg["content"]) > cfg.max_message_length:
+            raise ValueError("invalid message")
+    if messages[-1]["role"] != "user":
+        raise ValueError("last message is not a user message")
+    return messages
+
+
+def _validate_context(context: t.Any, cfg: SettingsAISummary) -> list[dict[str, str]]:
+    if not isinstance(context, list) or len(context) > cfg.max_context_items:
+        raise ValueError("invalid context")
+    for item in context:
+        if not isinstance(item, dict) or not item.keys() <= {"title", "url", "snippet"}:
+            raise ValueError("invalid context item")
+        for val in item.values():
+            if not isinstance(val, str) or len(val) > cfg.max_message_length:
+                raise ValueError("invalid context item")
+    return context
+
+
+def _validate_payload(payload: t.Any, cfg: SettingsAISummary) -> tuple[list[dict[str, str]], list[dict[str, str]]]:
+    """Validate the request body of the ``/ai_summary`` endpoint and return
+    the ``messages`` and ``context`` lists.  Raises a :py:obj:`ValueError` for
+    any malformed payload."""
+
+    if not isinstance(payload, dict):
+        raise ValueError("payload is not an object")
+    return _validate_messages(payload.get("messages"), cfg), _validate_context(payload.get("context", []), cfg)
+
+
+def ai_summary_view() -> flask.Response:
+    """Stream an AI generated answer for the messages in the request body,
+    response is NDJSON: ``{"delta": ..}`` lines followed by one final
+    ``{"done": true, ..}`` line."""
+
+    cfg: SettingsAISummary = get_setting("ai_summary")
+
+    if SXNGPlugin.id not in sxng_request.user_plugins:
+        return flask.Response(json.dumps({"error": "plugin is not enabled"}), status=403, mimetype="application/json")
+
+    try:
+        messages, context = _validate_payload(sxng_request.get_json(force=True, silent=True), cfg)
+    except ValueError as exc:
+        return _bad_request(str(exc))
+
+    server = _user_server(sxng_request, cfg)
+    if not _valid_server(server):
+        return _bad_request("no valid Ollama server configured")
+
+    model = str(sxng_request.preferences.get_value("ai_summary_model") or "").strip() or cfg.model
+    if not MODEL_NAME_REGEXP.fullmatch(model):
+        return _bad_request("no valid model configured")
+
+    ollama_payload = {
+        "model": model,
+        "messages": build_ollama_messages(cfg, messages, context),
+        "stream": True,
+        "keep_alive": cfg.keep_alive,
+    }
+
+    # open the upstream connection before streaming, a connection error is
+    # reported as HTTP 502 instead of a line in an already started stream
+    client = _get_client(server, cfg)
+    stream_ctx = client.stream("POST", "/api/chat", json=ollama_payload)
+    upstream = None
+    try:
+        upstream = stream_ctx.__enter__()  # pylint: disable=unnecessary-dunder-call
+        if upstream.status_code != 200:
+            stream_ctx.__exit__(None, None, None)
+            upstream = None
+    except httpx.HTTPError:
+        upstream = None
+    if upstream is None:
+        client.close()
+        return flask.Response(json.dumps({"error": "upstream error"}), status=502, mimetype="application/json")
+
+    # from here on nothing must be read from the request context, the
+    # generator runs after the request context has been torn down
+    log = logging.getLogger("searx.plugins.ai_summary")
+
+    def ndjson(obj: dict[str, t.Any]) -> bytes:
+        # the generator bypasses flask's response encoding (direct_passthrough)
+        return (json.dumps(obj) + "\n").encode()
+
+    def generate():
+        start = time.monotonic()
+        try:
+            for line in upstream.iter_lines():
+                if time.monotonic() - start > cfg.stream_timeout:
+                    yield ndjson({"done": True, "error": "timeout"})
+                    return
+                if not line.strip():
+                    continue
+                data = json.loads(line)
+                if data.get("done"):
+                    yield ndjson({"done": True, "model": model})
+                    return
+                delta = data.get("message", {}).get("content", "")
+                if delta:
+                    yield ndjson({"delta": delta})
+            yield ndjson({"done": True, "model": model})
+        except (httpx.HTTPError, ValueError) as exc:
+            log.warning("error while streaming from Ollama: %s", exc)
+            yield ndjson({"done": True, "error": "upstream error"})
+        finally:
+            stream_ctx.__exit__(None, None, None)
+            client.close()
+
+    return flask.Response(
+        generate(),
+        mimetype="application/x-ndjson",
+        headers={"Cache-Control": "no-store", "X-Accel-Buffering": "no"},
+        direct_passthrough=True,
+    )

searx/preferences.py

@@ -462,6 +462,20 @@ class Preferences:
                 locked="doi_resolver" in self.cfg.lock,
                 choices=DOI_RESOLVERS,
             ),
+            # empty values fall back to the administrator's defaults in the
+            # ai_summary: section (searx.ai_summary.SettingsAISummary)
+            'ai_summary_server': StringSetting(
+                "",
+                locked="ai_summary_server" in self.cfg.lock,
+            ),
+            'ai_summary_model': StringSetting(
+                "",
+                locked="ai_summary_model" in self.cfg.lock,
+            ),
+            'ai_summary_grounding': BooleanSetting(
+                get_setting("ai_summary").grounding,
+                locked="ai_summary_grounding" in self.cfg.lock,
+            ),
             'simple_style': EnumStringSetting(
                 get_setting("ui.theme_args.simple_style"),
                 locked="simple_style" in self.cfg.lock,

searx/result_types/__init__.py

@@ -19,6 +19,7 @@ __all__ = [
     "EngineResults",
     "AnswerSet",
     "Answer",
+    "AiSummary",
     "Translations",
     "WeatherAnswer",
     "Code",
@@ -32,7 +33,7 @@ import typing as t
 import abc
 
 from ._base import Result, MainResult, LegacyResult
-from .answer import AnswerSet, Answer, Translations, WeatherAnswer
+from .answer import AnswerSet, Answer, AiSummary, Translations, WeatherAnswer
 from .keyvalue import KeyValue
 from .code import Code
 from .paper import Paper
@@ -49,6 +50,7 @@ class ResultList(list[Result | LegacyResult], abc.ABC):
         implemented)."""
 
         Answer = Answer
+        AiSummary = AiSummary
         KeyValue = KeyValue
         Code = Code
         Paper = Paper

searx/result_types/answer.py

@@ -14,6 +14,10 @@ template.
    :members:
    :show-inheritance:
 
+.. autoclass:: AiSummary
+   :members:
+   :show-inheritance:
+
 .. autoclass:: Translations
    :members:
    :show-inheritance:
@@ -29,7 +33,7 @@ template.
 # pylint: disable=too-few-public-methods
 
 
-__all__ = ["AnswerSet", "Answer", "Translations", "WeatherAnswer"]
+__all__ = ["AnswerSet", "Answer", "AiSummary", "Translations", "WeatherAnswer"]
 
 from flask_babel import gettext
 import msgspec
@@ -92,6 +96,28 @@ class Answer(BaseAnswer, kw_only=True):
         return hash(self.answer)
 
 
+class AiSummary(BaseAnswer, kw_only=True):
+    """Placeholder for an AI generated summary of the search query
+    (:py:obj:`searx.plugins.ai_summary`).  The summary itself is fetched
+    asynchronously by the client after the result page has been rendered."""
+
+    # The answers of an AnswerSet are sorted by their template name; this
+    # template sorts before the other answer templates, the AI summary is
+    # therefore rendered first in the answer area.
+    template: str = "answer/ai_summary.html"
+
+    query: str
+    """The search query the summary is generated for."""
+
+    grounding: bool = False
+    """Ground the summary on the search results (user's preference)."""
+
+    def __hash__(self):
+        """The hash value of field *query* is the hash value of the
+        :py:obj:`AiSummary` object."""
+        return hash(self.query)
+
+
 class Translations(BaseAnswer, kw_only=True):
     """Answer type with a list of translations.
 

searx/settings.yml

@@ -257,6 +257,9 @@ plugins:
   searx.plugins.tracker_url_remover.SXNGPlugin:
     active: true
 
+  searx.plugins.ai_summary.SXNGPlugin:
+    active: false
+
 
 # Configuration of the "Hostnames plugin":
 #
@@ -284,6 +287,27 @@ plugins:
 # '(.*\.)?youtu\.be$': 'yt.example.com'
 #
 
+# Configuration of the "AI summary plugin", for more details see
+#   https://docs.searxng.org/admin/settings/settings_ai_summary.html
+#
+# ai_summary:
+#
+#   # Base URL of the Ollama server; without this URL the plugin is inactive.
+#   base_url: "http://127.0.0.1:11434"
+#
+#   # Default model; if empty, the first entry of models: is used.
+#   model: "llama3.2:3b"
+#
+#   # Models the user can select from in the preferences; if empty, the list
+#   # is requested from the Ollama server (GET /api/tags) at startup.
+#   models:
+#     - "llama3.2:3b"
+#     - "gemma3:4b"
+#
+#   # Ground summaries on the search results by default (users can still opt
+#   # in/out in their preferences).
+#   grounding: false
+
 
 categories_as_tabs:
   general:

searx/settings_defaults.py

@@ -13,6 +13,7 @@ from os.path import dirname, abspath
 import msgspec
 
 from typing_extensions import override
+from .ai_summary import SettingsAISummary
 from .brand import SettingsBrand
 from .sxng_locales import sxng_locales
 from ._settings import SettingsPref
@@ -267,6 +268,7 @@ SCHEMA: dict[str, t.Any] = {
         'networks': {},
     },
     'plugins': SettingsValue(dict, {}),
+    'ai_summary': SettingsAISummary,
     'categories_as_tabs': SettingsValue(dict, CATEGORIES_AS_TABS),
     'engines': SettingsValue(list, []),
     'doi_resolvers': {},

searx/templates/simple/answer/ai_summary.html

@@ -0,0 +1,19 @@
+<div id="ai_summary" class="ai-summary hide_if_nojs" {{- ' ' -}}
+     data-query="{{ answer.query }}" {{- ' ' -}}
+     data-grounding="{{ '1' if answer.grounding else '0' }}">{{- '' -}}
+  <div class="ai-summary-header">{{- '' -}}
+    <span class="ai-summary-title">{{ _('AI Summary') }}</span>{{- '' -}}
+    <span class="ai-summary-model"></span>{{- '' -}}
+  </div>{{- '' -}}
+  <div class="ai-summary-body collapsed">{{- '' -}}
+    <div class="ai-summary-answers" aria-live="polite"></div>{{- '' -}}
+  </div>{{- '' -}}
+  <button type="button" class="ai-summary-more invisible" {{- ' ' -}}
+          data-btn-text-collapsed="{{ _('More') }}" {{- ' ' -}}
+          data-btn-text-not-collapsed="{{ _('Less') }}">{{ _('More') }}</button>{{- '' -}}
+  <form class="ai-summary-followup invisible">{{- '' -}}
+    <input type="text" placeholder="{{ _('Ask a follow-up question') }}" autocomplete="off" maxlength="2048">{{- '' -}}
+    <button type="submit">{{ _('Ask') }}</button>{{- '' -}}
+  </form>{{- '' -}}
+  <p class="ai-summary-disclaimer">{{ _('Generated by AI — may contain mistakes.') }}</p>{{- '' -}}
+</div>

searx/templates/simple/preferences.html

@@ -249,6 +249,15 @@
     {%- endif -%}
     {{- tab_footer() -}}
 
+    {# tab: ai #}
+
+    {%- if plugins_storage | selectattr('preference_section', 'equalto', 'ai') | list -%}
+      {{- tab_header('maintab', 'ai', _('AI Summary')) -}}
+      {{- plugin_preferences('ai') -}}
+      {%- include 'simple/preferences/ai_summary.html' -%}
+      {{- tab_footer() -}}
+    {%- endif -%}
+
     {# tab: cookies #}
 
     {{- tab_header('maintab', 'cookies', _('Cookies')) -}}

searx/templates/simple/preferences/ai_summary.html

@@ -0,0 +1,59 @@
+{%- if 'ai_summary_server' not in locked_preferences -%}
+<fieldset>{{- '' -}}
+  <legend id="pref_ai_summary_server">{{- _('Ollama server URL') -}}</legend>{{- '' -}}
+  <div class="value">{{- '' -}}
+    <input name="ai_summary_server" aria-labelledby="pref_ai_summary_server" type="text"
+           autocomplete="off" spellcheck="false" autocorrect="off"
+           placeholder="{{ ai_summary_default_server or 'http://127.0.0.1:11434' }}"
+           value="{{ preferences.get_value('ai_summary_server') }}">{{- '' -}}
+  </div>{{- '' -}}
+  <div class="description">
+    {{- _('URL of the Ollama server that generates the summaries, e.g. http://192.168.1.10:11434.') -}}
+    {{- ' ' -}}
+    {%- if ai_summary_default_server -%}
+      {{- _('Leave empty to use the default of this instance.') -}}
+    {%- endif -%}
+  </div>{{- '' -}}
+</fieldset>{{- '' -}}
+{%- endif -%}
+{%- if 'ai_summary_model' not in locked_preferences -%}
+<fieldset>{{- '' -}}
+  <legend id="pref_ai_summary_model">{{- _('AI summary model') -}}</legend>{{- '' -}}
+  <div class="value">{{- '' -}}
+    <input name="ai_summary_model" aria-labelledby="pref_ai_summary_model" type="text"
+           autocomplete="off" spellcheck="false" autocorrect="off" list="ai_summary_model_list"
+           placeholder="{{ ai_summary_default_model or 'llama3.2:3b' }}"
+           value="{{ preferences.get_value('ai_summary_model') }}">{{- '' -}}
+    <datalist id="ai_summary_model_list">
+      {%- for model in ai_summary_models -%}
+        <option value="{{ model }}"></option>
+      {%- endfor -%}
+    </datalist>{{- '' -}}
+  </div>{{- '' -}}
+  <div class="description">
+    {{- _('Name of the model used to generate the summaries, e.g. llama3.2:3b or gemma3:4b.') -}}
+    {{- ' ' -}}
+    {%- if ai_summary_default_model -%}
+      {{- _('Leave empty to use the default of this instance.') -}}
+    {%- endif -%}
+  </div>{{- '' -}}
+</fieldset>{{- '' -}}
+{%- endif -%}
+{%- if 'ai_summary_grounding' not in locked_preferences -%}
+<fieldset>{{- '' -}}
+  <legend id="pref_ai_summary_grounding">{{ _('Ground AI summary on search results') }}</legend>{{- '' -}}
+  <p class="value">{{- '' -}}
+    <input type="checkbox" {{- ' ' -}}
+           name="ai_summary_grounding" {{- ' ' -}}
+           aria-labelledby="pref_ai_summary_grounding" {{- ' ' -}}
+           class="checkbox-onoff" {{- ' ' -}}
+           {%- if preferences.get_value('ai_summary_grounding') -%}
+             checked
+           {%- endif -%}{{- ' ' -}}
+           >{{- '' -}}
+  </p>{{- '' -}}
+  <div class="description">
+    {{- _('Send the top search results along with the query, the model answers from this context instead of its own knowledge.  Answers are more accurate and more current, but generating them is slower and needs more memory (VRAM) on the Ollama server.') -}}
+  </div>{{- '' -}}
+</fieldset>{{- '' -}}
+{%- endif -%}

searx/webapp.py

@@ -93,8 +93,10 @@ from searx.preferences import (
     ClientPref,
     ValidationException,
 )
+import searx.ai_summary
 import searx.answerers
 import searx.plugins
+import searx.plugins.ai_summary
 
 
 from searx.metrics import get_engines_stats, get_engine_errors, get_reliabilities, histogram, counter, openmetrics
@@ -328,6 +330,8 @@ def get_translations():
         'Source': gettext('Source'),
         # infinite scroll
         'error_loading_next_page': gettext('Error loading the next page'),
+        # AI summary
+        'error_loading_ai_summary': gettext('Error loading the AI summary'),
     }
 
 
@@ -981,11 +985,17 @@ def preferences():
         preferences_url_params = sxng_request.preferences.get_as_url_params(),
         locked_preferences = get_setting("preferences").lock,
         doi_resolvers = get_setting("doi_resolvers", {}),
+        ai_summary_models = searx.ai_summary.model_choices(),
+        ai_summary_default_model = get_setting("ai_summary").model,
+        ai_summary_default_server = get_setting("ai_summary").base_url,
         # fmt: on
     )
 
 
 app.add_url_rule('/favicon_proxy', methods=['GET'], endpoint="favicon_proxy", view_func=favicons.favicon_proxy)
+app.add_url_rule(
+    '/ai_summary', methods=['POST'], endpoint="ai_summary", view_func=searx.plugins.ai_summary.ai_summary_view
+)
 
 
 @app.route('/image_proxy', methods=['GET'])

tests/unit/test_plugin_ai_summary.py

@@ -0,0 +1,271 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+# pylint: disable=missing-module-docstring,missing-class-docstring,invalid-name,protected-access
+# pylint: disable=too-many-public-methods
+
+import json
+from contextlib import contextmanager
+
+from mock import Mock
+
+import searx.ai_summary
+import searx.plugins
+import searx.plugins.ai_summary
+import searx.preferences
+
+from searx.extended_types import sxng_request
+from searx.result_types import AiSummary
+
+from tests import SearxTestCase
+from .test_plugins import get_search_mock
+
+PLUGIN_FQN = "searx.plugins.ai_summary.SXNGPlugin"
+BASE_URL = "http://127.0.0.1:11434"
+MODEL = "test-model"
+
+
+def ollama_stream_mock(lines: list[dict], status_code: int = 200) -> Mock:
+    """A mock httpx client whose ``stream()`` context manager yields the given
+    (Ollama) NDJSON lines."""
+
+    upstream = Mock(status_code=status_code)
+    upstream.iter_lines.return_value = iter([json.dumps(line) for line in lines])
+
+    @contextmanager
+    def stream(*_args, **_kwargs):
+        yield upstream
+
+    client = Mock()
+    client.stream = stream
+    return client
+
+
+class PluginAISummaryInit(SearxTestCase):
+
+    def test_active_without_base_url(self):
+        # the Ollama server can be configured by the user in the preferences,
+        # the plugin stays active without an admin configured default
+        self.setattr4test(searx.ai_summary, "MODELS", [])
+
+        storage = searx.plugins.PluginStorage()
+        storage.load_settings({PLUGIN_FQN: {"active": True}})
+        storage.init(self.app)
+
+        self.assertEqual(1, len(storage))
+        self.assertEqual([], searx.ai_summary.MODELS)
+
+    def test_model_list_from_settings(self):
+        cfg = searx.get_setting("ai_summary")
+        self.setattr4test(cfg, "base_url", BASE_URL)
+        self.setattr4test(cfg, "model", MODEL)
+        self.setattr4test(cfg, "models", [MODEL, "other-model"])
+        self.setattr4test(searx.ai_summary, "MODELS", [])
+
+        storage = searx.plugins.PluginStorage()
+        storage.load_settings({PLUGIN_FQN: {"active": True}})
+        storage.init(self.app)
+
+        self.assertEqual(1, len(storage))
+        self.assertEqual([MODEL, "other-model"], searx.ai_summary.MODELS)
+
+
+class PluginAISummary(SearxTestCase):
+
+    def setUp(self):
+        super().setUp()
+
+        cfg = searx.get_setting("ai_summary")
+        self.setattr4test(cfg, "base_url", BASE_URL)
+        self.setattr4test(cfg, "model", MODEL)
+        self.setattr4test(cfg, "models", [MODEL])
+        self.setattr4test(searx.ai_summary, "MODELS", [])
+
+        self.storage = searx.plugins.PluginStorage()
+        self.storage.load_settings({PLUGIN_FQN: {"active": True}})
+        self.storage.init(self.app)
+
+        # the endpoint checks request.user_plugins (built in webapp.pre_request
+        # from the global plugin storage) -- enable the plugin like a browser
+        # with saved preferences does
+        self.client.set_cookie("disabled_plugins", "")
+        self.client.set_cookie("enabled_plugins", "ai_summary")
+
+        self.pref = searx.preferences.Preferences(["simple"], ["general"], {}, self.storage)
+        self.pref.parse_dict({"locale": "en"})
+
+    def mock_upstream(self, client_mock: Mock):
+        self.setattr4test(searx.plugins.ai_summary, "_get_client", lambda *_args, **_kwargs: client_mock)
+
+    def do_post_search(self, query, **kwargs) -> Mock:
+        kwargs.setdefault("categories", ["general"])
+        search = get_search_mock(query, user_plugins=["ai_summary"], **kwargs)
+        self.storage.post_search(sxng_request, search)
+        return search
+
+    def test_placeholder_answer_is_added(self):
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+
+            search = self.do_post_search("what is the best searx fork")
+            answer = AiSummary(query="what is the best searx fork", grounding=False)
+            self.assertIn(answer, search.result_container.answers)
+
+    def test_placeholder_carries_grounding_pref(self):
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+            self.pref.parse_dict({"ai_summary_grounding": "1"})
+
+            search = self.do_post_search("lorem ipsum")
+            answer = list(search.result_container.answers)[0]
+            self.assertTrue(answer.grounding)
+
+    def test_grounding_default_from_settings(self):
+        self.setattr4test(searx.get_setting("ai_summary"), "grounding", True)
+        pref = searx.preferences.Preferences(["simple"], ["general"], {}, self.storage)
+
+        with self.app.test_request_context():
+            sxng_request.preferences = pref
+            search = self.do_post_search("lorem ipsum")
+            answer = list(search.result_container.answers)[0]
+            self.assertTrue(answer.grounding)
+
+    def test_skip_pageno(self):
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+            search = self.do_post_search("lorem ipsum", pageno=2)
+            self.assertEqual(list(search.result_container.answers), [])
+
+    def test_skip_non_html_format(self):
+        with self.app.test_request_context("/search", method="POST", data={"format": "json"}):
+            sxng_request.preferences = self.pref
+            search = self.do_post_search("lorem ipsum")
+            self.assertEqual(list(search.result_container.answers), [])
+
+    def test_skip_non_general_category(self):
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+            search = self.do_post_search("lorem ipsum", categories=["images"])
+            self.assertEqual(list(search.result_container.answers), [])
+
+    def test_skip_infobox(self):
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+            search = get_search_mock("lorem ipsum", user_plugins=["ai_summary"], categories=["general"])
+            search.result_container.infoboxes.append(Mock())
+            self.storage.post_search(sxng_request, search)
+            self.assertEqual(list(search.result_container.answers), [])
+
+    def test_skip_instant_answer(self):
+        # e.g. the "ddg definitions" engine adds wikipedia abstracts as Answer
+        from searx.result_types import Answer  # pylint: disable=import-outside-toplevel
+
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+            search = get_search_mock("lorem ipsum", user_plugins=["ai_summary"], categories=["general"])
+            engine_answer = Answer(answer="Lorem ipsum is placeholder text.  More at Wikipedia")
+            search.result_container.answers.add(engine_answer)
+            self.storage.post_search(sxng_request, search)
+            self.assertEqual(list(search.result_container.answers), [engine_answer])
+
+    def test_skip_without_any_server(self):
+        self.setattr4test(searx.get_setting("ai_summary"), "base_url", "")
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+            search = self.do_post_search("lorem ipsum")
+            self.assertEqual(list(search.result_container.answers), [])
+
+    def test_placeholder_with_user_server_only(self):
+        self.setattr4test(searx.get_setting("ai_summary"), "base_url", "")
+        with self.app.test_request_context():
+            sxng_request.preferences = self.pref
+            self.pref.parse_dict({"ai_summary_server": "http://192.168.1.10:11434"})
+            search = self.do_post_search("lorem ipsum")
+            self.assertEqual(len(list(search.result_container.answers)), 1)
+
+    def test_endpoint_forbidden_when_disabled(self):
+        self.client.set_cookie("disabled_plugins", "ai_summary")
+        self.client.set_cookie("enabled_plugins", "")
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 403)
+
+    def test_endpoint_bad_request(self):
+        for body in [
+            None,
+            {},
+            {"messages": []},
+            {"messages": [{"role": "system", "content": "hi"}]},
+            {"messages": [{"role": "user", "content": "hi"}, {"role": "assistant", "content": "ho"}]},
+            {"messages": [{"role": "user", "content": "x" * 5000}]},
+            {"messages": [{"role": "user", "content": "hi"}], "context": [{"unknown_key": "x"}]},
+            {"messages": [{"role": "user", "content": "hi"}] * 13},
+        ]:
+            res = self.client.post("/ai_summary", json=body)
+            self.assertEqual(res.status_code, 400, body)
+
+    def test_endpoint_invalid_server_pref(self):
+        self.client.set_cookie("ai_summary_server", "ftp://example.org")
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 400)
+
+    def test_endpoint_invalid_model_pref(self):
+        self.client.set_cookie("ai_summary_model", "bad model name!")
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 400)
+
+    def test_endpoint_no_server_configured(self):
+        self.setattr4test(searx.get_setting("ai_summary"), "base_url", "")
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 400)
+
+    def test_endpoint_streams_ndjson(self):
+        self.mock_upstream(
+            ollama_stream_mock(
+                [
+                    {"message": {"content": "Hello "}, "done": False},
+                    {"message": {"content": "world"}, "done": False},
+                    {"done": True},
+                ]
+            )
+        )
+
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 200)
+        self.assertEqual(res.headers["Content-Type"], "application/x-ndjson")
+
+        lines = [json.loads(line) for line in res.data.decode().splitlines() if line]
+        self.assertEqual(lines[0], {"delta": "Hello "})
+        self.assertEqual(lines[1], {"delta": "world"})
+        self.assertEqual(lines[2], {"done": True, "model": MODEL})
+
+    def test_endpoint_model_pref_wins(self):
+        self.mock_upstream(ollama_stream_mock([{"done": True}]))
+        self.client.set_cookie("ai_summary_model", "my-own-model:7b")
+
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        lines = [json.loads(line) for line in res.data.decode().splitlines() if line]
+        self.assertEqual(lines[-1], {"done": True, "model": "my-own-model:7b"})
+
+    def test_endpoint_upstream_error(self):
+        self.mock_upstream(ollama_stream_mock([], status_code=500))
+
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 502)
+
+    def test_endpoint_error_while_streaming(self):
+        upstream = Mock(status_code=200)
+        upstream.iter_lines.return_value = iter(
+            [json.dumps({"message": {"content": "Hello"}, "done": False}), "this is not json"]
+        )
+
+        @contextmanager
+        def stream(*_args, **_kwargs):
+            yield upstream
+
+        client_mock = Mock()
+        client_mock.stream = stream
+        self.mock_upstream(client_mock)
+
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 200)
+        lines = [json.loads(line) for line in res.data.decode().splitlines() if line]
+        self.assertEqual(lines[0], {"delta": "Hello"})
+        self.assertEqual(lines[1], {"done": True, "error": "upstream error"})