jasonwitty
4abb7dba67
Talk to the LLM server via GET /v1/models and POST /v1/chat/completions (SSE) instead of Ollama's native API. Any OpenAI compatible server now works (Ollama, vLLM, llama.cpp, LM Studio, Hugging Face TGI, ...); Ollama serves this API natively, existing setups keep working unchanged. The Ollama specific keep_alive option is dropped, the ai_summary.grounding setting is added as instance wide default of the grounding preference. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
314 lines
13 KiB
Python
314 lines
13 KiB
Python
# 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) LLM server that
|
|
implements the `OpenAI chat completions API`_ -- e.g. `Ollama`_, vLLM,
|
|
llama.cpp, LM Studio or Hugging Face TGI.
|
|
|
|
The LLM 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 re-emits the SSE token stream of the
|
|
LLM server's ``/v1/chat/completions`` 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 LLM server are sent directly (not via
|
|
:py:obj:`searx.network`), an outgoing proxy configuration is deliberately not
|
|
applied to reach an LLM 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/
|
|
.. _OpenAI chat completions API: https://platform.openai.com/docs/api-reference/chat
|
|
.. _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_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_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 LLM server at ``base_url``."""
|
|
# 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,
|
|
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 LLM 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 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) 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 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
|
|
client = _get_client(server, cfg)
|
|
stream_ctx = client.stream("POST", "/chat/completions", json=chat_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:
|
|
# 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,
|
|
)
|