[feat] plugin: AI summary of search results from a self-hosted LLM

Adds an optional plugin that shows a generated summary above the search
results, similar to the answer boxes in Brave and Google.  The text is
produced by an LLM server the administrator runs, reached over the
OpenAI chat completions API, so queries never leave the operator's own
network.

The summary is grounded on the top search results rather than the
model's training data.  It is generated asynchronously: post_search adds
an empty placeholder and returns, and the browser fills it from the
/ai_summary endpoint, which streams the answer as NDJSON.  No summary is
generated beyond page one, outside the general category, for non-HTML
formats, or when an engine already answered with an infobox or an
instant answer.

The client side follows the existing plugin pattern: one file in
client/simple/src/js/plugin/, one conditional load in router.ts, one
LESS import.  No build configuration changes and no new dependencies.

The plugin is not activated by default.  Instance defaults live in an
ai_summary: section; the server, model, API key and grounding are user
preferences, and all four can be locked.

Signed-off-by: Jason Witty <jasonpwitty+github@proton.me>

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 +1,135 @@
+# 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_chat_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 LLM server (e.g. ``http://127.0.0.1:11434``
+    for Ollama).  Any server that implements the OpenAI chat completions API
+    works (Ollama, vLLM, llama.cpp, LM Studio, Hugging Face TGI, ...).  Users
+    can set their own server URL in the preferences (``ai_summary_server``)
+    unless that preference is locked."""
+
+    api_key: str = ""
+    """Optional API key of the LLM server in
+    :py:obj:`SettingsAISummary.base_url`, sent in an ``Authorization: Bearer``
+    header.  Needed by servers that require authentication, e.g. vLLM or
+    llama.cpp started with ``--api-key``, or an LLM server behind an
+    authenticating reverse proxy.
+
+    This key is **only** sent to :py:obj:`SettingsAISummary.base_url`: a user
+    who points the ``ai_summary_server`` preference at a server of their own
+    gets no ``Authorization`` header from it, so the key can't be captured by
+    a third party.  For their own server, users configure their own key in the
+    ``ai_summary_api_key`` preference (see
+    :py:obj:`searx.plugins.ai_summary._server_api_key`)."""
+
+    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 LLM server (``GET /v1/models``)."""
+
+    grounding: bool = True
+    """Default of the ``ai_summary_grounding`` user preference: ground the
+    summary on the search results.  Grounded summaries are more accurate and
+    more current at a moderate extra cost (the search results are sent along
+    with the query, so the prompt is longer).  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 LLM 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."""
+
+    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_chat_messages(
+    cfg: SettingsAISummary,
+    messages: list[dict[str, str]],
+    context: list[dict[str, str]] | None = None,
+) -> list[dict[str, str]]:
+    """Build the message list for the chat completions request 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 +1,386 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+"""Implementation of the AI Summary plugin, which shows a generated answer above
+the search results.  The answer comes from an LLM server that implements the
+`OpenAI chat completions API`_ (Ollama, vLLM, llama.cpp, LM Studio, Hugging Face
+TGI, ...) and that the administrator runs.
+
+- :ref:`ai_summary plugin` describes the design and the request flow.
+- :ref:`settings ai_summary` describes how to configure it.
+
+This module holds the plugin itself and the ``/ai_summary`` endpoint
+(:py:obj:`ai_summary_view`, registered in :py:obj:`searx.webapp`).  The endpoint
+streams the answer to the browser, so that the result page is never delayed by
+the LLM; :py:obj:`SXNGPlugin.post_search` only adds an empty
+:py:obj:`searx.result_types.AiSummary` placeholder for the client to fill.
+
+Settings of the ``ai_summary:`` section are defined in
+:py:obj:`searx.ai_summary.SettingsAISummary`.
+
+.. _OpenAI chat completions API: https://platform.openai.com/docs/api-reference/chat
+"""
+
+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_chat_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 names differ per provider: "gemma3:4b" (Ollama), "bedrock/anthropic.
+# claude-3-5-sonnet" (a gateway's routing prefix), "gemini-1.5-pro@001" (a
+# pinned version).  The pattern accepts those and rejects anything that could
+# change the meaning of the request body it is placed into.
+MODEL_NAME_REGEXP = re.compile(r"[A-Za-z0-9._:/@-]{1,128}")
+
+log = logging.getLogger("searx.plugins.ai_summary")
+
+UPSTREAM_RETRIES = 1
+"""How often a request to the LLM server is repeated when the server does not
+answer in time.
+
+An idle LLM server unloads the model, and loads it again on the next request --
+which can take longer than :py:obj:`read_timeout
+<searx.ai_summary.SettingsAISummary.read_timeout>`, because no byte of the
+response is sent while the model is loading.  The request that runs into this
+is also the request that starts the load, so repeating it usually succeeds."""
+
+
+def _get_client(base_url: str, cfg: SettingsAISummary, api_key: str = "") -> httpx.Client:
+    """HTTP client for one request to the LLM server at ``base_url``.  The
+    ``api_key`` (if any) is sent in an ``Authorization: Bearer`` header, see
+    :py:obj:`_server_api_key`."""
+    # the OpenAI API paths are prefixed with /v1, unless the base URL already
+    # points into an API prefix
+    base_url = base_url.rstrip("/")
+    if not base_url.endswith("/v1"):
+        base_url += "/v1"
+    return httpx.Client(
+        base_url=base_url,
+        headers={"Authorization": f"Bearer {api_key}"} if api_key else None,
+        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 _server_id(url: str) -> tuple[str, str, int, str] | None:
+    """Identity of an LLM server URL (scheme, host, port, path) for comparing
+    two URLs, or ``None`` if the URL is unusable.  The ``/v1`` API prefix is
+    not part of the identity, :py:obj:`_get_client` appends it when missing."""
+    try:
+        parsed = urlparse(url)
+    except ValueError:
+        return None
+    # .hostname (not .netloc) drops the userinfo, so that a server URL like
+    # http://llm.example.org@untrusted.example.org/ is identified by the host
+    # the request is actually sent to (untrusted.example.org)
+    if parsed.scheme not in ("http", "https") or not parsed.hostname:
+        return None
+    port = parsed.port or (443 if parsed.scheme == "https" else 80)
+    path = parsed.path.rstrip("/")
+    if path.endswith("/v1"):
+        path = path[: -len("/v1")].rstrip("/")
+    return (parsed.scheme, parsed.hostname.lower(), port, path)
+
+
+def _server_api_key(cfg: SettingsAISummary, server: str, user_api_key: str = "") -> str:
+    """The API key to send to ``server``:
+
+    - the administrator's :py:obj:`cfg.api_key
+      <searx.ai_summary.SettingsAISummary.api_key>` if ``server`` *is* the
+      administrator's server (:py:obj:`cfg.base_url
+      <searx.ai_summary.SettingsAISummary.base_url>`),
+    - otherwise the user's own ``ai_summary_api_key`` preference, which belongs
+      to the server in the user's own ``ai_summary_server`` preference.
+
+    The administrator's key is never sent to a server a user configured --
+    that would hand every user of the instance a way to capture it."""
+    server_id = _server_id(server)
+    if server_id is not None and server_id == _server_id(cfg.base_url):
+        return cfg.api_key
+    return user_api_key
+
+
+def _user_server(request: "SXNG_Request", cfg: SettingsAISummary) -> str:
+    """The LLM server URL for this request: the user's ``ai_summary_server``
+    preference, or the administrator's default."""
+    server = str(request.preferences.get_value("ai_summary_server") or "").strip()
+    # Credentials are stripped from a user's server URL: httpx turns them into
+    # an Authorization header, and a user should not be able to make SearXNG
+    # send a header of their choosing to a host of their choosing.  An
+    # administrator can still use credentials in the configured base_url (e.g.
+    # an LLM server behind basic auth).
+    if server and "@" in urlparse(server).netloc:
+        server = ""
+    return server 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 a local LLM 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 LLM server (``GET
+        /v1/models``).  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, cfg.api_key) as client:
+                resp = client.get("/models")
+                resp.raise_for_status()
+                models = [model["id"] for model in resp.json().get("data", [])]
+        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 LLM 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 _open_upstream(client: httpx.Client, payload: dict[str, t.Any]) -> tuple[t.Any, t.Any]:
+    """Start the streaming completion on the LLM server.
+
+    Returns the (already entered) stream context and the response, or
+    ``(None, None)`` if no usable response was received.
+
+    The request is repeated (:py:obj:`UPSTREAM_RETRIES`) when the server did not
+    answer in time, and when it answered ``5xx`` -- both mean *not right now*,
+    and the most common reason is a model that is still being loaded.  A ``4xx``
+    is not repeated: a wrong API key or an unknown model name does not become
+    right when asked twice."""
+
+    for attempt in range(UPSTREAM_RETRIES + 1):
+        stream_ctx = client.stream("POST", "/chat/completions", json=payload)
+        try:
+            resp = stream_ctx.__enter__()  # pylint: disable=unnecessary-dunder-call
+        except httpx.TransportError as exc:
+            if attempt < UPSTREAM_RETRIES:
+                log.debug("LLM server did not answer (%s), asking again", exc)
+                continue
+            log.warning("LLM server did not answer: %s", exc)
+            return None, None
+        except httpx.HTTPError as exc:
+            log.warning("request to the LLM server failed: %s", exc)
+            return None, None
+
+        if resp.status_code == 200:
+            return stream_ctx, resp
+
+        # the body of an error response is short and usually names the cause,
+        # e.g. an unknown model; without it a misconfiguration is invisible
+        detail = ""
+        try:
+            resp.read()
+            detail = resp.text.strip()[:200]
+        except (httpx.HTTPError, UnicodeDecodeError):  # pragma: no cover
+            pass
+        stream_ctx.__exit__(None, None, None)
+
+        # 5xx is the server saying it is not able to answer *right now* -- a
+        # model still loading, a gateway with no upstream yet.  4xx is the
+        # server saying the request is wrong, which a second one would be too.
+        if resp.status_code >= 500 and attempt < UPSTREAM_RETRIES:
+            log.debug("LLM server replied HTTP %s (%s), asking again", resp.status_code, detail)
+            continue
+        log.warning("LLM server responded with HTTP %s %s", resp.status_code, detail)
+        return None, None
+
+    return None, None  # pragma: no cover - the loop always returns
+
+
+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 LLM 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")
+
+    chat_payload = {
+        "model": model,
+        "messages": build_chat_messages(cfg, messages, context),
+        "stream": True,
+    }
+
+    # open the upstream connection before streaming, a connection error is
+    # reported as HTTP 502 instead of a line in an already started stream
+    user_api_key = str(sxng_request.preferences.get_value("ai_summary_api_key") or "").strip()
+    client = _get_client(server, cfg, _server_api_key(cfg, server, user_api_key))
+    stream_ctx, upstream = _open_upstream(client, chat_payload)
+    if stream_ctx is None or 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
+
+    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:
+            # the upstream is a SSE stream: "data: {..}" lines, terminated by
+            # a "data: [DONE]" line
+            for line in upstream.iter_lines():
+                if time.monotonic() - start > cfg.stream_timeout:
+                    yield ndjson({"done": True, "error": "timeout"})
+                    return
+                line = line.strip()
+                if not line or line.startswith(":") or not line.startswith("data:"):
+                    continue
+                payload = line[len("data:") :].strip()
+                if payload == "[DONE]":
+                    break
+                data = json.loads(payload)
+                choices = data.get("choices") or [{}]
+                delta = choices[0].get("delta", {}).get("content") or ""
+                if delta:
+                    yield ndjson({"delta": delta})
+            yield ndjson({"done": True, "model": model})
+        except (httpx.HTTPError, ValueError) as exc:
+            log.warning("error while streaming from the LLM server: %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

@@ -49,10 +49,14 @@ class ValidationException(Exception):
 class Setting:
     """Base class of user settings"""
 
-    def __init__(self, default_value: t.Any, locked: bool = False):
+    def __init__(self, default_value: t.Any, locked: bool = False, secret: bool = False):
         super().__init__()
         self.value: t.Any = default_value
         self.locked: bool = locked
+        self.secret: bool = secret
+        """The value is a credential: it is not included in the preferences URL
+        (:py:obj:`Preferences.get_as_url_params`), which users copy around to
+        transfer or share their preferences."""
 
     def parse(self, data: str):
         """Parse ``data`` and store the result at ``self.value``
@@ -462,6 +466,27 @@ 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_api_key': StringSetting(
+                "",
+                locked="ai_summary_api_key" in self.cfg.lock,
+                # a user's API key is only sent to a server the user configured
+                # themselves, and it is never part of the preferences URL
+                secret=True,
+            ),
+            '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,
@@ -498,7 +523,7 @@ class Preferences:
         """Return preferences as URL parameters"""
         settings_kv = {}
         for k, v in self.key_value_settings.items():
-            if v.locked:
+            if v.locked or v.secret:
                 continue
             if isinstance(v, MultipleChoiceSetting):
                 settings_kv[k] = ','.join(v.get_value())

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

@@ -256,6 +256,9 @@ plugins:
   searx.plugins.tracker_url_remover.SXNGPlugin:
     active: true
 
+  searx.plugins.ai_summary.SXNGPlugin:
+    active: false
+
 
 # Configuration of the "Hostnames plugin":
 #
@@ -283,6 +286,34 @@ 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 an OpenAI compatible LLM server (Ollama, vLLM, LM Studio,
+#   # llama.cpp, Hugging Face TGI, ...), used as the default for the
+#   # ai_summary_server preference.
+#   base_url: "http://127.0.0.1:11434"
+#
+#   # API key of the server above, if it requires authentication (e.g. vLLM or
+#   # llama.cpp with --api-key).  Sent as "Authorization: Bearer" and only to
+#   # base_url, never to a server configured by a user in the preferences.
+#   api_key: ""
+#
+#   # Default model; if empty, the first entry of models: is used.
+#   model: "llama3.2:3b"
+#
+#   # Models suggested to the user in the preferences; if empty, the list
+#   # is requested from the LLM server (GET /v1/models) 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: true
+
 
 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/static/themes/simple/chunk/BuurKv-k.min.js

@@ -1,2 +0,0 @@
-var e=class{id;constructor(e){this.id=e,queueMicrotask(()=>this.invoke())}async invoke(){try{console.debug(`[PLUGIN] ${this.id}: Running...`);let e=await this.run();if(!e)return;console.debug(`[PLUGIN] ${this.id}: Running post-exec...`),await this.post(e)}catch(e){console.error(`[PLUGIN] ${this.id}:`,e)}finally{console.debug(`[PLUGIN] ${this.id}: Done.`)}}};export{e as t};
-//# sourceMappingURL=BuurKv-k.min.js.map

searx/static/themes/simple/chunk/BuurKv-k.min.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"BuurKv-k.min.js","names":[],"sources":["../../../../../client/simple/src/js/Plugin.ts"],"sourcesContent":["// SPDX-License-Identifier: AGPL-3.0-or-later\n\n/**\n * Base class for client-side plugins.\n *\n * @remarks\n * Handle conditional loading of the plugin in:\n *\n *   - client/simple/src/js/router.ts\n *\n * @abstract\n */\nexport abstract class Plugin {\n  /**\n   * Plugin name.\n   */\n  protected readonly id: string;\n\n  /**\n   * @remarks\n   * Don't hold references of this instance outside the class.\n   */\n  protected constructor(id: string) {\n    this.id = id;\n\n    queueMicrotask(() => this.invoke());\n  }\n\n  private async invoke(): Promise<void> {\n    try {\n      console.debug(`[PLUGIN] ${this.id}: Running...`);\n      const result = await this.run();\n      if (!result) return;\n\n      console.debug(`[PLUGIN] ${this.id}: Running post-exec...`);\n      // @ts-expect-error\n      void (await this.post(result as NonNullable<Awaited<ReturnType<this[\"run\"]>>>));\n    } catch (error) {\n      console.error(`[PLUGIN] ${this.id}:`, error);\n    } finally {\n      console.debug(`[PLUGIN] ${this.id}: Done.`);\n    }\n  }\n\n  /**\n   * Plugin goes here.\n   *\n   * @remarks\n   * The plugin is already loaded at this point. If you wish to execute\n   * conditions to exit early, consider moving the logic to:\n   *\n   *   - client/simple/src/js/router.ts\n   *\n   * ...to avoid unnecessarily loading this plugin on the client.\n   */\n  protected abstract run(): Promise<unknown>;\n\n  /**\n   * Post-execution hook.\n   *\n   * @remarks\n   * The hook is only executed if `#run()` returns a truthy value.\n   */\n  // @ts-expect-error\n  protected abstract post(result: NonNullable<Awaited<ReturnType<this[\"run\"]>>>): Promise<void>;\n}\n"],"mappings":"AAYA,IAAsB,EAAtB,KAA6B,CAI3B,GAMA,YAAsB,EAAY,CAChC,KAAK,GAAK,EAEV,mBAAqB,KAAK,OAAO,CAAC,CACpC,CAEA,MAAc,QAAwB,CACpC,GAAI,CACF,QAAQ,MAAM,YAAY,KAAK,GAAG,aAAa,EAC/C,IAAM,EAAS,MAAM,KAAK,IAAI,EAC9B,GAAI,CAAC,EAAQ,OAEb,QAAQ,MAAM,YAAY,KAAK,GAAG,uBAAuB,EAEzD,MAAY,KAAK,KAAK,CAAuD,CAC/E,OAAS,EAAO,CACd,QAAQ,MAAM,YAAY,KAAK,GAAG,GAAI,CAAK,CAC7C,QAAU,CACR,QAAQ,MAAM,YAAY,KAAK,GAAG,QAAQ,CAC5C,CACF,CAuBF"}

searx/static/themes/simple/chunk/D3mcqWOe.min.js

@@ -1,2 +0,0 @@
-import{t as e}from"./BuurKv-k.min.js";import{i as t,t as n}from"../sxng-core.min.js";import{t as r}from"./DK4yUVpy.min.js";import{t as i}from"./DcK-mo-Y.min.js";var a=class extends e{constructor(){super(`infiniteScroll`)}async run(){let e=i(`results`).classList.contains(`only_template_images`),a=`article.result:last-child`,o=document.createElement(`div`);o.className=`loader`;let s=async i=>{let a=document.querySelector(`#search`);r(a);let s=document.querySelector(`#pagination form.next_page`);r(s);let c=a.getAttribute(`action`);if(!c)throw Error(`Form action not defined`);let l=document.querySelector(`#pagination`);r(l),l.replaceChildren(o);try{let t=await(await n(`POST`,c,{body:new FormData(s)})).text();if(!t)return;let r=new DOMParser().parseFromString(t,`text/html`),a=r.querySelectorAll(`#urls article`),o=r.querySelector(`#pagination`);document.querySelector(`#pagination`)?.remove();let l=document.querySelector(`#urls`);if(!l)throw Error(`URLs element not found`);a.length>0&&!e&&l.appendChild(document.createElement(`hr`)),l.append(...a),o&&(document.querySelector(`#results`)?.appendChild(o),i())}catch(e){console.error(`Error loading next page:`,e);let n=Object.assign(document.createElement(`div`),{textContent:t.translations?.error_loading_next_page??`Error loading next page`,className:`dialog-error`});n.setAttribute(`role`,`alert`),document.querySelector(`#pagination`)?.replaceChildren(n)}},c=new IntersectionObserver(async e=>{let[t]=e;t?.isIntersecting&&(c.unobserve(t.target),await s(()=>{let e=document.querySelector(a);e&&c.observe(e)}))},{rootMargin:`320px`}),l=document.querySelector(a);l&&c.observe(l)}async post(){}};export{a as default};
-//# sourceMappingURL=D3mcqWOe.min.js.map

searx/static/themes/simple/chunk/DBBLNmaq.min.js

@@ -0,0 +1,4 @@
+import{a as e,i as t}from"../sxng-core.min.js";import{t as n}from"./DK4yUVpy.min.js";var r=5,i=12,a=300,o=1e3,s=class s extends e{messages=[];context=[];controller;constructor(){super(`ai_summary`)}async run(){let e=document.getElementById(`ai_summary`);if(e)try{let{query:t}=e.dataset;if(!t)return;e.dataset.grounding===`1`&&(this.context=s.collectContext()),this.wireControls(e),this.messages.push({role:`user`,content:t}),await this.exchange(e)}catch(t){s.showError(e,t)}}async post(){}static collectContext(){let e=[];for(let t of document.querySelectorAll(`#urls article.result`)){if(e.length>=r)break;let n=t.querySelector(`h3 a`);n&&e.push({title:(n.textContent??``).trim().slice(0,a),url:n.href,snippet:(t.querySelector(`.content`)?.textContent??``).trim().slice(0,o)})}return e}wireControls(e){let t=e.querySelector(`.ai-summary-body`),r=e.querySelector(`.ai-summary-more`),a=e.querySelector(`.ai-summary-followup`);n(t),n(r),n(a),r.addEventListener(`click`,()=>{let e=t.classList.toggle(`collapsed`);r.textContent=e?r.dataset.btnTextCollapsed??``:r.dataset.btnTextNotCollapsed??``,a.classList.toggle(`invisible`,e)}),a.addEventListener(`submit`,t=>{t.preventDefault();let r=a.querySelector(`input`);n(r);let o=r.value.trim();if(!o||this.controller)return;r.value=``,this.messages.push({role:`user`,content:o}),this.messages.splice(0,this.messages.length-(i-1));let s=Object.assign(document.createElement(`p`),{textContent:o,className:`ai-summary-question`});e.querySelector(`.ai-summary-answers`)?.append(s),this.exchange(e)})}async exchange(e){let t=e.querySelector(`.ai-summary-answers`);n(t);let r=Object.assign(document.createElement(`p`),{className:`ai-summary-content typing`});t.append(r);let i=new AbortController;this.controller=i;let a=``,o=t=>{if(!t.trim())return;let n=JSON.parse(t);if(n.error)throw Error(n.error);if(n.delta&&(a+=n.delta,r.textContent=a,this.updateMoreButton(e)),n.done&&n.model){let t=e.querySelector(`.ai-summary-model`);t&&(t.textContent=n.model)}};try{let e=await fetch(`./ai_summary`,{method:`POST`,headers:{"Content-Type":`application/json`},body:JSON.stringify({messages:this.messages,context:this.context}),signal:i.signal});if(!e.ok)throw Error(`HTTP ${e.status}`);if(e.body){let t=e.body.getReader(),n=new TextDecoder,r=``;for(;;){let{done:e,value:i}=await t.read();if(e)break;r+=n.decode(i,{stream:!0});let a=r.split(`
+`);r=a.pop()??``;for(let e of a)o(e)}o(r)}else{let t=await e.text();for(let e of t.split(`
+`))o(e)}this.messages.push({role:`assistant`,content:a})}catch(t){s.showError(e,t)}finally{r.classList.remove(`typing`),this.controller=void 0,this.updateMoreButton(e)}}static showError(e,n){console.error(`Error loading AI summary:`,n);let r=t.translations?.error_loading_ai_summary??`Error loading the AI summary`,i=n instanceof Error&&n.message?` (${n.message})`:``,a=Object.assign(document.createElement(`div`),{textContent:`${r}${i}`,className:`dialog-error`});a.setAttribute(`role`,`alert`),(e.querySelector(`.ai-summary-answers`)??e).append(a)}updateMoreButton(e){let t=e.querySelector(`.ai-summary-body`),n=e.querySelector(`.ai-summary-more`);t&&n&&(!t.classList.contains(`collapsed`)||t.scrollHeight>t.clientHeight+4)&&n.classList.remove(`invisible`)}};export{s as default};
+//# sourceMappingURL=DBBLNmaq.min.js.map

searx/static/themes/simple/chunk/DpvWr1cn.min.js

@@ -0,0 +1,2 @@
+import{a as e,i as t,t as n}from"../sxng-core.min.js";import{t as r}from"./DK4yUVpy.min.js";import{t as i}from"./DcK-mo-Y.min.js";var a=class extends e{constructor(){super(`infiniteScroll`)}async run(){let e=i(`results`).classList.contains(`only_template_images`),a=`article.result:last-child`,o=document.createElement(`div`);o.className=`loader`;let s=async i=>{let a=document.querySelector(`#search`);r(a);let s=document.querySelector(`#pagination form.next_page`);r(s);let c=a.getAttribute(`action`);if(!c)throw Error(`Form action not defined`);let l=document.querySelector(`#pagination`);r(l),l.replaceChildren(o);try{let t=await(await n(`POST`,c,{body:new FormData(s)})).text();if(!t)return;let r=new DOMParser().parseFromString(t,`text/html`),a=r.querySelectorAll(`#urls article`),o=r.querySelector(`#pagination`);document.querySelector(`#pagination`)?.remove();let l=document.querySelector(`#urls`);if(!l)throw Error(`URLs element not found`);a.length>0&&!e&&l.appendChild(document.createElement(`hr`)),l.append(...a),o&&(document.querySelector(`#results`)?.appendChild(o),i())}catch(e){console.error(`Error loading next page:`,e);let n=Object.assign(document.createElement(`div`),{textContent:t.translations?.error_loading_next_page??`Error loading next page`,className:`dialog-error`});n.setAttribute(`role`,`alert`),document.querySelector(`#pagination`)?.replaceChildren(n)}},c=new IntersectionObserver(async e=>{let[t]=e;t?.isIntersecting&&(c.unobserve(t.target),await s(()=>{let e=document.querySelector(a);e&&c.observe(e)}))},{rootMargin:`320px`}),l=document.querySelector(a);l&&c.observe(l)}async post(){}};export{a as default};
+//# sourceMappingURL=DpvWr1cn.min.js.map

searx/static/themes/simple/manifest.json

@@ -1,8 +1,4 @@
 {
-  "_BuurKv-k.min.js": {
-    "file": "chunk/BuurKv-k.min.js",
-    "name": "plugin"
-  },
   "_DK4yUVpy.min.js": {
     "file": "chunk/DK4yUVpy.min.js",
     "name": "assertelement"
@@ -23,6 +19,7 @@
       "src/js/plugin/MapView.ts",
       "src/js/plugin/InfiniteScroll.ts",
       "src/js/plugin/Calculator.ts",
+      "src/js/plugin/AiSummary.ts",
       "src/js/main/keyboard.ts",
       "src/js/main/search.ts",
       "src/js/main/autocomplete.ts",
@@ -80,35 +77,44 @@
       "_DcK-mo-Y.min.js"
     ]
   },
+  "src/js/plugin/AiSummary.ts": {
+    "file": "chunk/DBBLNmaq.min.js",
+    "name": "aisummary",
+    "src": "src/js/plugin/AiSummary.ts",
+    "isDynamicEntry": true,
+    "imports": [
+      "src/js/index.ts",
+      "_DK4yUVpy.min.js"
+    ]
+  },
   "src/js/plugin/Calculator.ts": {
-    "file": "chunk/BfLIj3Of.min.js",
+    "file": "chunk/C8c7HJzp.min.js",
     "name": "calculator",
     "src": "src/js/plugin/Calculator.ts",
     "isDynamicEntry": true,
     "imports": [
-      "_BuurKv-k.min.js",
+      "src/js/index.ts",
       "_DcK-mo-Y.min.js"
     ]
   },
   "src/js/plugin/InfiniteScroll.ts": {
-    "file": "chunk/D3mcqWOe.min.js",
+    "file": "chunk/DpvWr1cn.min.js",
     "name": "infinitescroll",
     "src": "src/js/plugin/InfiniteScroll.ts",
     "isDynamicEntry": true,
     "imports": [
-      "_BuurKv-k.min.js",
       "src/js/index.ts",
       "_DK4yUVpy.min.js",
       "_DcK-mo-Y.min.js"
     ]
   },
   "src/js/plugin/MapView.ts": {
-    "file": "chunk/Bc8fcwWx.min.js",
+    "file": "chunk/BnnvKC7b.min.js",
     "name": "mapview",
     "src": "src/js/plugin/MapView.ts",
     "isDynamicEntry": true,
     "imports": [
-      "_BuurKv-k.min.js"
+      "src/js/index.ts"
     ],
     "css": [
       "sxng-mapview.min.css"

searx/static/themes/simple/sxng-core.min.js

@@ -1,3 +1,3 @@
-const __vite__mapDeps=(i,m=__vite__mapDeps,d=(m.f||(m.f=["./chunk/Bc8fcwWx.min.js","./chunk/BuurKv-k.min.js","./sxng-mapview.min.css","./chunk/D3mcqWOe.min.js","./chunk/DK4yUVpy.min.js","./chunk/DcK-mo-Y.min.js","./chunk/BfLIj3Of.min.js","./chunk/C93hSkpT.min.js","./chunk/5Ako-qGW.min.js","./chunk/C21EfLAC.min.js","./chunk/od7pNHfk.min.js","./chunk/e2-9fzwE.min.js"])))=>i.map(i=>d[i]);
-var e={index:`index`,results:`results`,preferences:`preferences`,unknown:`unknown`},t={closeDetail:void 0,scrollPageToSelected:void 0,selectImage:void 0,selectNext:void 0,selectPrevious:void 0},n=()=>{let t=document.querySelector(`meta[name="endpoint"]`)?.getAttribute(`content`);return t&&t in e?t:e.unknown},r=()=>{let e=document.querySelector(`script[client_settings]`)?.getAttribute(`client_settings`);if(!e)return{};try{return JSON.parse(atob(e))}catch(e){return console.error(`Failed to load client_settings:`,e),{}}},i=async(e,t,n)=>{let r=new AbortController,i=setTimeout(()=>r.abort(),n?.timeout??3e4),a=await fetch(t,{body:n?.body,method:e,signal:r.signal}).finally(()=>clearTimeout(i));if(!a.ok)throw Error(a.statusText);return a},a=(e,t,n,r)=>{if(typeof t!=`string`){t.addEventListener(e,n,r);return}document.addEventListener(e,e=>{for(let r of e.composedPath())if(r instanceof HTMLElement&&r.matches(t)){try{n.call(r,e)}catch(e){console.error(e)}break}},r)},o=(e,t)=>{for(let e of t?.on??[])if(!e)return;document.readyState===`loading`?a(`DOMContentLoaded`,document,e,{once:!0}):e()},s=n(),c=r(),l=(e,t)=>{u(t)&&e()},u=e=>{switch(e.on){case`global`:return!0;case`endpoint`:return!!e.where.includes(s)}},d=`modulepreload`,f=function(e,t){return new URL(e,t).href},p={},m=function(e,t,n){let r=Promise.resolve();if(t&&t.length>0){let e=document.getElementsByTagName(`link`),i=document.querySelector(`meta[property=csp-nonce]`),a=i?.nonce||i?.getAttribute(`nonce`);function o(e){return Promise.all(e.map(e=>Promise.resolve(e).then(e=>({status:`fulfilled`,value:e}),e=>({status:`rejected`,reason:e}))))}function s(e){return import.meta.resolve?import.meta.resolve(e):new URL(e,import.meta.url).href}r=o(t.map(t=>{if(t=f(t,n),t=s(t),t in p)return;p[t]=!0;let r=t.endsWith(`.css`);for(let n=e.length-1;n>=0;n--){let i=e[n];if(i.href===t&&(!r||i.rel===`stylesheet`))return}let i=document.createElement(`link`);if(i.rel=r?`stylesheet`:d,r||(i.as=`script`),i.crossOrigin=``,i.href=t,a&&i.setAttribute(`nonce`,a),document.head.appendChild(i),r)return new Promise((e,n)=>{i.addEventListener(`load`,e),i.addEventListener(`error`,()=>n(Error(`Unable to preload CSS for ${t}`)))})}))}function i(e){let t=new Event(`vite:preloadError`,{cancelable:!0});if(t.payload=e,window.dispatchEvent(t),!t.defaultPrevented)throw e}return r.then(t=>{for(let e of t||[])e.status===`rejected`&&i(e.reason);return e().catch(i)})};o(()=>{document.documentElement.classList.remove(`no-js`),document.documentElement.classList.add(`js`),a(`click`,`.close`,function(){this.parentNode?.classList.add(`invisible`)}),a(`click`,`.searxng_init_map`,async function(t){t.preventDefault(),this.classList.remove(`searxng_init_map`),l(()=>m(async()=>{let{default:e}=await import(`./chunk/Bc8fcwWx.min.js`);return{default:e}},__vite__mapDeps([0,1,2]),import.meta.url).then(({default:e})=>new e(this)),{on:`endpoint`,where:[e.results]})}),c.plugins?.includes(`infiniteScroll`)&&l(()=>m(async()=>{let{default:e}=await import(`./chunk/D3mcqWOe.min.js`);return{default:e}},__vite__mapDeps([3,1,4,5]),import.meta.url).then(({default:e})=>new e),{on:`endpoint`,where:[e.results]}),c.plugins?.includes(`calculator`)&&l(()=>m(async()=>{let{default:e}=await import(`./chunk/BfLIj3Of.min.js`);return{default:e}},__vite__mapDeps([6,1,5,4]),import.meta.url).then(({default:e})=>new e),{on:`endpoint`,where:[e.results]})}),o(()=>{m(()=>import(`./chunk/C93hSkpT.min.js`),__vite__mapDeps([7,4]),import.meta.url),m(()=>import(`./chunk/5Ako-qGW.min.js`),__vite__mapDeps([8,5,4]),import.meta.url),c.autocomplete&&m(()=>import(`./chunk/C21EfLAC.min.js`),__vite__mapDeps([9,4]),import.meta.url)},{on:[s===e.index]}),o(()=>{m(()=>import(`./chunk/C93hSkpT.min.js`),__vite__mapDeps([7,4]),import.meta.url),m(()=>import(`./chunk/od7pNHfk.min.js`),__vite__mapDeps([10,4]),import.meta.url),m(()=>import(`./chunk/5Ako-qGW.min.js`),__vite__mapDeps([8,5,4]),import.meta.url),c.autocomplete&&m(()=>import(`./chunk/C21EfLAC.min.js`),__vite__mapDeps([9,4]),import.meta.url)},{on:[s===e.results]}),o(()=>{m(()=>import(`./chunk/e2-9fzwE.min.js`),__vite__mapDeps([11,4]),import.meta.url)},{on:[s===e.preferences]});export{c as i,a as n,t as r,i as t};
+const __vite__mapDeps=(i,m=__vite__mapDeps,d=(m.f||(m.f=["./chunk/BnnvKC7b.min.js","./sxng-mapview.min.css","./chunk/DpvWr1cn.min.js","./chunk/DK4yUVpy.min.js","./chunk/DcK-mo-Y.min.js","./chunk/C8c7HJzp.min.js","./chunk/DBBLNmaq.min.js","./chunk/C93hSkpT.min.js","./chunk/5Ako-qGW.min.js","./chunk/C21EfLAC.min.js","./chunk/od7pNHfk.min.js","./chunk/e2-9fzwE.min.js"])))=>i.map(i=>d[i]);
+var e=class{id;constructor(e){this.id=e,queueMicrotask(()=>this.invoke())}async invoke(){try{console.debug(`[PLUGIN] ${this.id}: Running...`);let e=await this.run();if(!e)return;console.debug(`[PLUGIN] ${this.id}: Running post-exec...`),await this.post(e)}catch(e){console.error(`[PLUGIN] ${this.id}:`,e)}finally{console.debug(`[PLUGIN] ${this.id}: Done.`)}}},t={index:`index`,results:`results`,preferences:`preferences`,unknown:`unknown`},n={closeDetail:void 0,scrollPageToSelected:void 0,selectImage:void 0,selectNext:void 0,selectPrevious:void 0},r=()=>{let e=document.querySelector(`meta[name="endpoint"]`)?.getAttribute(`content`);return e&&e in t?e:t.unknown},i=()=>{let e=document.querySelector(`script[client_settings]`)?.getAttribute(`client_settings`);if(!e)return{};try{return JSON.parse(atob(e))}catch(e){return console.error(`Failed to load client_settings:`,e),{}}},a=async(e,t,n)=>{let r=new AbortController,i=setTimeout(()=>r.abort(),n?.timeout??3e4),a=await fetch(t,{body:n?.body,method:e,signal:r.signal}).finally(()=>clearTimeout(i));if(!a.ok)throw Error(a.statusText);return a},o=(e,t,n,r)=>{if(typeof t!=`string`){t.addEventListener(e,n,r);return}document.addEventListener(e,e=>{for(let r of e.composedPath())if(r instanceof HTMLElement&&r.matches(t)){try{n.call(r,e)}catch(e){console.error(e)}break}},r)},s=(e,t)=>{for(let e of t?.on??[])if(!e)return;document.readyState===`loading`?o(`DOMContentLoaded`,document,e,{once:!0}):e()},c=r(),l=i(),u=(e,t)=>{d(t)&&e()},d=e=>{switch(e.on){case`global`:return!0;case`endpoint`:return!!e.where.includes(c)}},f=`modulepreload`,p=function(e,t){return new URL(e,t).href},m={},h=function(e,t,n){let r=Promise.resolve();if(t&&t.length>0){let e=document.getElementsByTagName(`link`),i=document.querySelector(`meta[property=csp-nonce]`),a=i?.nonce||i?.getAttribute(`nonce`);function o(e){return Promise.all(e.map(e=>Promise.resolve(e).then(e=>({status:`fulfilled`,value:e}),e=>({status:`rejected`,reason:e}))))}function s(e){return import.meta.resolve?import.meta.resolve(e):new URL(e,import.meta.url).href}r=o(t.map(t=>{if(t=p(t,n),t=s(t),t in m)return;m[t]=!0;let r=t.endsWith(`.css`);for(let n=e.length-1;n>=0;n--){let i=e[n];if(i.href===t&&(!r||i.rel===`stylesheet`))return}let i=document.createElement(`link`);if(i.rel=r?`stylesheet`:f,r||(i.as=`script`),i.crossOrigin=``,i.href=t,a&&i.setAttribute(`nonce`,a),document.head.appendChild(i),r)return new Promise((e,n)=>{i.addEventListener(`load`,e),i.addEventListener(`error`,()=>n(Error(`Unable to preload CSS for ${t}`)))})}))}function i(e){let t=new Event(`vite:preloadError`,{cancelable:!0});if(t.payload=e,window.dispatchEvent(t),!t.defaultPrevented)throw e}return r.then(t=>{for(let e of t||[])e.status===`rejected`&&i(e.reason);return e().catch(i)})};s(()=>{document.documentElement.classList.remove(`no-js`),document.documentElement.classList.add(`js`),o(`click`,`.close`,function(){this.parentNode?.classList.add(`invisible`)}),o(`click`,`.searxng_init_map`,async function(e){e.preventDefault(),this.classList.remove(`searxng_init_map`),u(()=>h(async()=>{let{default:e}=await import(`./chunk/BnnvKC7b.min.js`);return{default:e}},__vite__mapDeps([0,1]),import.meta.url).then(({default:e})=>new e(this)),{on:`endpoint`,where:[t.results]})}),l.plugins?.includes(`infiniteScroll`)&&u(()=>h(async()=>{let{default:e}=await import(`./chunk/DpvWr1cn.min.js`);return{default:e}},__vite__mapDeps([2,3,4]),import.meta.url).then(({default:e})=>new e),{on:`endpoint`,where:[t.results]}),l.plugins?.includes(`calculator`)&&u(()=>h(async()=>{let{default:e}=await import(`./chunk/C8c7HJzp.min.js`);return{default:e}},__vite__mapDeps([5,4,3]),import.meta.url).then(({default:e})=>new e),{on:`endpoint`,where:[t.results]}),l.plugins?.includes(`ai_summary`)&&u(()=>h(async()=>{let{default:e}=await import(`./chunk/DBBLNmaq.min.js`);return{default:e}},__vite__mapDeps([6,3]),import.meta.url).then(({default:e})=>new e),{on:`endpoint`,where:[t.results]})}),s(()=>{h(()=>import(`./chunk/C93hSkpT.min.js`),__vite__mapDeps([7,3]),import.meta.url),h(()=>import(`./chunk/5Ako-qGW.min.js`),__vite__mapDeps([8,4,3]),import.meta.url),l.autocomplete&&h(()=>import(`./chunk/C21EfLAC.min.js`),__vite__mapDeps([9,3]),import.meta.url)},{on:[c===t.index]}),s(()=>{h(()=>import(`./chunk/C93hSkpT.min.js`),__vite__mapDeps([7,3]),import.meta.url),h(()=>import(`./chunk/od7pNHfk.min.js`),__vite__mapDeps([10,3]),import.meta.url),h(()=>import(`./chunk/5Ako-qGW.min.js`),__vite__mapDeps([8,4,3]),import.meta.url),l.autocomplete&&h(()=>import(`./chunk/C21EfLAC.min.js`),__vite__mapDeps([9,3]),import.meta.url)},{on:[c===t.results]}),s(()=>{h(()=>import(`./chunk/e2-9fzwE.min.js`),__vite__mapDeps([11,3]),import.meta.url)},{on:[c===t.preferences]});export{e as a,l as i,o as n,n as r,a as t};
 //# sourceMappingURL=sxng-core.min.js.map

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 +249,18 @@
     {%- endif -%}
     {{- tab_footer() -}}
 
+    {# tab: ai #}
+
+    {#- the tab is only shown when the administrator activated the plugin in
+        settings.yml, an instance without an AI summary has no AI tab -#}
+    {%- if 'ai_summary' in plugins_active_by_default
+           and 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 +1,72 @@
+{%- if 'ai_summary_server' not in locked_preferences -%}
+<fieldset>{{- '' -}}
+  <legend id="pref_ai_summary_server">{{- _('AI 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 OpenAI compatible LLM server that generates the summaries (e.g. Ollama, LM Studio, vLLM), 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_api_key' not in locked_preferences -%}
+<fieldset>{{- '' -}}
+  <legend id="pref_ai_summary_api_key">{{- _('AI server API key') -}}</legend>{{- '' -}}
+  <div class="value">{{- '' -}}
+    <input name="ai_summary_api_key" aria-labelledby="pref_ai_summary_api_key" type="password"
+           autocomplete="off" spellcheck="false" autocorrect="off"
+           value="{{ preferences.get_value('ai_summary_api_key') }}">{{- '' -}}
+  </div>{{- '' -}}
+  <div class="description">
+    {{- _('Optional, only needed if your server requires authentication.') -}}
+  </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'),
     }
 
 
@@ -975,16 +979,25 @@ def preferences():
         shortcuts = {y: x for x, y in engine_shortcuts.items()},
         themes = themes,
         plugins_storage = searx.plugins.STORAGE.info,
+        # plugins the administrator activated in settings.yml; a plugin that is
+        # not activated does not get a preferences tab of its own
+        plugins_active_by_default = {plg.id for plg in searx.plugins.STORAGE if plg.active},
         current_doi_resolver = get_doi_resolver(),
         allowed_plugins = allowed_plugins,
         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'])