LLM AI integration

changedetectionio/blueprint/settings/templates/settings.html

@@ -1,7 +1,7 @@
 {% extends 'base.html' %}
 
 {% block content %}
-{% from '_helpers.html' import render_field, render_checkbox_field, render_button, render_time_schedule_form %}
+{% from '_helpers.html' import render_field, render_checkbox_field, render_simple_field, render_button, render_time_schedule_form %}
 {% from '_common_fields.html' import render_common_settings_form %}
 <script>
     const notification_base_url="{{url_for('ui.ui_notification.ajax_callback_send_notification_test', mode="global-settings")}}";
@@ -23,6 +23,7 @@
             <li class="tab"><a href="#fetching">Fetching</a></li>
             <li class="tab"><a href="#filters">Global Filters</a></li>
             <li class="tab"><a href="#ui-options">UI Options</a></li>
+            <li class="tab"><a href="#ai-options"><i data-feather="aperture" style="width: 14px; height: 14px; margin-right: 4px;"></i> AI</a></li>
             <li class="tab"><a href="#api">API</a></li>
             <li class="tab"><a href="#timedate">Time &amp Date</a></li>
             <li class="tab"><a href="#proxies">CAPTCHA &amp; Proxies</a></li>
@@ -262,6 +263,24 @@ <h4>Chrome Extension</h4>
                 </div>
 
             </div>
+            <div class="tab-pane-inner" id="ai-options">
+                <p><strong>New:</strong> click here (link to changedetection.io tutorial page) find out how to setup and example</p>
+                <br>
+                key fields should be some password type field so you can see its set but doesnt contain the key on view and doesnt lose it on save<br>
+
+                <div class="pure-control-group inline-radio">
+                    {{ render_simple_field(form.application.form.ai.form.LLM_backend) }}
+                    <span class="pure-form-message-inline">Preferred LLM connection</span>
+                </div>
+                <div class="pure-control-group">
+                    {{ render_checkbox_field(form.application.form.ai.form.API_keys.form.openai) }}
+                    <span class="pure-form-message-inline">Go here to read more about OpenAI integration</span>
+                </div>
+                <div class="pure-control-group">
+                    {{ render_checkbox_field(form.application.form.ai.form.API_keys.form.gemini) }}
+                    <span class="pure-form-message-inline">Go here to read more about Google Gemini integration</span>
+                </div>
+            </div>
             <div class="tab-pane-inner" id="proxies">
                 <div id="recommended-proxy">
                     <div>

changedetectionio/blueprint/tags/templates/edit-tag.html

@@ -25,7 +25,7 @@
     <div class="tabs collapsable">
         <ul>
             <li class="tab" id=""><a href="#general">General</a></li>
-            <li class="tab"><a href="#filters-and-triggers">Filters &amp; Triggers</a></li>
+            <li class="tab"><a href="#filters-and-triggers">AI, Filters &amp; Triggers</a></li>
             {% if extra_tab_content %}
             <li class="tab"><a href="#extras_tab">{{ extra_tab_content }}</a></li>
             {% endif %}

changedetectionio/blueprint/ui/edit.py

@@ -312,8 +312,27 @@ def watch_get_preview_rendered(uuid):
         '''For when viewing the "preview" of the rendered text from inside of Edit'''
         from flask import jsonify
         from changedetectionio.processors.text_json_diff import prepare_filter_prevew
-        result = prepare_filter_prevew(watch_uuid=uuid, form_data=request.form, datastore=datastore)
-        return jsonify(result)
+
+        watch = datastore.data["watching"].get(uuid)
+
+        if not watch:
+            return jsonify({
+                "error": "Watch not found",
+                "code": 400
+            }), 400
+
+        if not watch.history_n:
+            return jsonify({
+                "error": "Watch has empty history, at least one fetch of the page is required.",
+                "code": 400
+            }), 400
+        #
+        try:
+            result = prepare_filter_prevew(watch_uuid=uuid, form_data=request.form, datastore=datastore)
+            return jsonify(result)
+        except Exception as e:
+            return abort(500, str(e))
+
 
     @edit_blueprint.route("/highlight_submit_ignore_url", methods=['POST'])
     @login_optionally_required

changedetectionio/blueprint/ui/views.py

@@ -212,7 +212,14 @@ def form_quick_watch_add():
 
         add_paused = request.form.get('edit_and_watch_submit_button') != None
         processor = request.form.get('processor', 'text_json_diff')
-        new_uuid = datastore.add_watch(url=url, tag=request.form.get('tags').strip(), extras={'paused': add_paused, 'processor': processor})
+        extras = {'paused': add_paused, 'processor': processor}
+
+        LLM_prompt = request.form.get('LLM_prompt', '').strip()
+        if LLM_prompt:
+            extras['LLM_prompt'] = LLM_prompt
+            extras['LLM_send_type'] = request.form.get('LLM_send_type', 'text')
+
+        new_uuid = datastore.add_watch(url=url, tag=request.form.get('tags').strip(), extras=extras)
 
         if new_uuid:
             if add_paused:

changedetectionio/blueprint/watchlist/templates/watch-overview.html

@@ -5,12 +5,7 @@
 <script src="{{url_for('static_content', group='js', filename='watch-overview.js')}}" defer></script>
 <script>let nowtimeserver={{ now_time_server }};</script>
 <script>let favicon_baseURL="{{ url_for('static_content', group='favicon', filename="PLACEHOLDER")}}";</script>
-<script>
-// Initialize Feather icons after the page loads
-document.addEventListener('DOMContentLoaded', function() {
-    feather.replace();
-});
-</script>
+
 <style>
 .checking-now .last-checked {
     background-image: linear-gradient(to bottom, transparent 0%, rgba(0,0,0,0.05) 40%, rgba(0,0,0,0.1) 100%);
@@ -31,8 +26,12 @@
                     {{ render_nolabel_field(form.edit_and_watch_submit_button, title="Edit first then Watch") }}
             </div>
             <div id="watch-group-tag">
+                <i data-feather="tag" style="width: 14px; height: 14px; stroke: white; margin-right: 4px;"></i>
                {{ render_field(form.tags, value=active_tag.title if active_tag_uuid else '', placeholder="Watch group / tag", class="transparent-field") }}
             </div>
+
+            {%- include 'edit/llm_prompt.html' -%}
+
             <div id="quick-watch-processor-type">
                 {{ render_simple_field(form.processor) }}
             </div>

changedetectionio/forms.py

@@ -55,6 +55,18 @@
 default_method = 'GET'
 allow_simplehost = not strtobool(os.getenv('BLOCK_SIMPLEHOSTS', 'False'))
 
+LLM_example_texts = ['Tell me simply "Price, In stock"',
+                     'Give me a list of all products for sale in this text',
+                     'Tell me simply "Yes" "No" or "Maybe" if you think the weather outlook is good for a 4-day small camping trip',
+                     'Look at this restaurant menu and only give me list of meals you think are good for type 2 diabetics, if nothing is found just say "nothing"',
+                     ]
+
+LLM_send_type_choices = [('text', 'Plain text after filters'),
+                         ('above_fold_text', 'Text above the fold'),
+                         ('Screenshot', 'Screenshot / Selection'),
+                         ('HTML', 'HTML Source')
+                         ]
+
 class StringListField(StringField):
     widget = widgets.TextArea()
 
@@ -515,18 +527,23 @@ def __call__(self, form, field):
 
 class quickWatchForm(Form):
     from . import processors
+    import random
 
     url = fields.URLField('URL', validators=[validateURL()])
     tags = StringTagUUID('Group tag', [validators.Optional()])
     watch_submit_button = SubmitField('Watch', render_kw={"class": "pure-button pure-button-primary"})
     processor = RadioField(u'Processor', choices=processors.available_processors(), default="text_json_diff")
+    LLM_prompt = TextAreaField(u'AI Prompt', [validators.Optional()], render_kw={"placeholder": f'Example, "{random.choice(LLM_example_texts)}"'})
+    LLM_send_type = RadioField(u'LLM Send', choices=LLM_send_type_choices, default="text")
+
     edit_and_watch_submit_button = SubmitField('Edit > Watch', render_kw={"class": "pure-button pure-button-primary"})
 
 
 
 # Common to a single watch and the global settings
 class commonSettingsForm(Form):
     from . import processors
+    import random
 
     def __init__(self, formdata=None, obj=None, prefix="", data=None, meta=None, **kwargs):
         super().__init__(formdata, obj, prefix, data, meta, **kwargs)
@@ -544,6 +561,8 @@ def __init__(self, formdata=None, obj=None, prefix="", data=None, meta=None, **k
     timezone = StringField("Timezone for watch schedule", render_kw={"list": "timezones"}, validators=[validateTimeZoneName()])
     webdriver_delay = IntegerField('Wait seconds before extracting text', validators=[validators.Optional(), validators.NumberRange(min=1, message="Should contain one or more seconds")])
 
+    LLM_prompt = TextAreaField(u'AI Prompt', [validators.Optional()], render_kw={"placeholder": f'Example, "{random.choice(LLM_example_texts)}"'})
+    LLM_send_type = RadioField(u'LLM Send', choices=LLM_send_type_choices, default="text")
 
 class importForm(Form):
     from . import processors
@@ -742,6 +761,29 @@ class globalSettingsApplicationUIForm(Form):
     socket_io_enabled = BooleanField('Realtime UI Updates Enabled', default=True, validators=[validators.Optional()])
     favicons_enabled = BooleanField('Favicons Enabled', default=True, validators=[validators.Optional()])
 
+class globalSettingsApplicationAIKeysForm(Form):
+
+    openai = StringField('OpenAI Key',
+                           validators=[validators.Optional()],
+                           render_kw={"placeholder": 'xxxxxxxxx'}
+                           )
+    gemini = StringField('Google Gemini Key',
+                           validators=[validators.Optional()],
+                           render_kw={"placeholder": 'ooooooooo'}
+                           )
+
+class globalSettingsApplicationAIForm(Form):
+
+    #@todo use only configured types?
+    LLM_backend = RadioField(u'LLM Backend',
+                               choices=[('openai', 'Open AI'), ('gemini', 'Gemini')],
+                               default="text")
+
+    # So that we can pass this to our LLM/__init__.py as a keys dict
+    API_keys = FormField(globalSettingsApplicationAIKeysForm)
+
+
+
 # datastore.data['settings']['application']..
 class globalSettingsApplicationForm(commonSettingsForm):
 
@@ -774,6 +816,8 @@ class globalSettingsApplicationForm(commonSettingsForm):
                                                                                                      message="Should contain zero or more attempts")])
     ui = FormField(globalSettingsApplicationUIForm)
 
+    ai = FormField(globalSettingsApplicationAIForm)
+
 
 class globalSettingsForm(Form):
     # Define these as FormFields/"sub forms", this way it matches the JSON storage

changedetectionio/model/App.py

@@ -65,6 +65,10 @@ class model(dict):
                         'socket_io_enabled': True,
                         'favicons_enabled': True
                     },
+                    'ai': {
+                        'openai_key': None,
+                        'gemini_key': None
+                    }
                 }
             }
         }

changedetectionio/model/__init__.py

@@ -38,6 +38,9 @@ def __init__(self, *arg, **kw):
             'last_error': False,
             'last_notification_error': None,
             'last_viewed': 0,  # history key value of the last viewed via the [diff] link
+            'LLM_prompt': None,
+            'LLM_send_type': None,
+            'LLM_backend': None,
             'method': 'GET',
             'notification_alert_count': 0,
             'notification_body': None,

changedetectionio/processors/LLM/__init__.py

@@ -0,0 +1,64 @@
+import importlib
+from langchain_core.messages import SystemMessage, HumanMessage
+
+SYSTEM_MESSAGE = (
+    "You are a text analyser who will attempt to give the most concise information "
+    "to the request, the information should be returned in a way that if I ask you again "
+    "I should get the same answer if the outcome is the same. The goal is to cut down "
+    "or reduce the text changes from you when i ask the same question about similar content "
+    "Always list items in exactly the same order and wording as found in the source text. "
+)
+
+
+class LLM_integrate:
+    PROVIDER_MAP = {
+        "openai": ("langchain_openai", "ChatOpenAI"),
+        "azure": ("langchain_community.chat_models", "AzureChatOpenAI"),
+        "gemini": ("langchain_google_genai", "ChatGoogleGenerativeAI")
+    }
+
+    def __init__(self, api_keys: dict):
+        """
+        api_keys = {
+            "openai": "sk-xxx",
+            "azure": "AZURE_KEY",
+            "gemini": "GEMINI_KEY"
+        }
+        """
+        self.api_keys = api_keys
+
+    def run(self, provider: str, model: str, message: str):
+        module_name, class_name = self.PROVIDER_MAP[provider]
+
+        # Import the class dynamically
+        module = importlib.import_module(module_name)
+        LLMClass = getattr(module, class_name)
+
+        # Create the LLM object
+        llm_kwargs = {}
+        if provider == "openai":
+            llm_kwargs = dict(api_key=self.api_keys.get("openai", ''),
+                              model=model,
+                              # https://api.python.langchain.com/en/latest/chat_models/langchain_openai.chat_models.base.ChatOpenAI.html#langchain_openai.chat_models.base.ChatOpenAI.temperature
+                              temperature=0 # most deterministic,
+                              )
+        elif provider == "azure":
+            llm_kwargs = dict(
+                api_key=self.api_keys["azure"],
+                azure_endpoint="https://<your-endpoint>.openai.azure.com",
+                deployment_name=model
+            )
+        elif provider == "gemini":
+            llm_kwargs = dict(api_key=self.api_keys.get("gemini"), model=model)
+
+        llm = LLMClass(**llm_kwargs)
+
+        # Build your messages
+        messages = [
+            SystemMessage(content=SYSTEM_MESSAGE),
+            HumanMessage(content=message)
+        ]
+
+        # Run the model asynchronously
+        result = llm.invoke(messages)
+        return result.content

changedetectionio/processors/__init__.py

@@ -1,5 +1,6 @@
 from abc import abstractmethod
 from changedetectionio.content_fetchers.base import Fetcher
+from changedetectionio.processors.LLM import LLM_integrate
 from changedetectionio.strtobool import strtobool
 from copy import deepcopy
 from loguru import logger

changedetectionio/processors/text_json_diff/processor.py

@@ -7,7 +7,7 @@
 import urllib3
 
 from changedetectionio.conditions import execute_ruleset_against_all_plugins
-from changedetectionio.processors import difference_detection_processor
+from changedetectionio.processors import difference_detection_processor, LLM_integrate
 from changedetectionio.html_tools import PERL_STYLE_REGEX, cdata_in_document_to_text, TRANSLATE_WHITESPACE_TABLE
 from changedetectionio import html_tools, content_fetchers
 from changedetectionio.blueprint.price_data_follower import PRICE_DATA_TRACK_ACCEPT, PRICE_DATA_TRACK_REJECT
@@ -293,6 +293,30 @@ def run_changedetection(self, watch):
             # we end up with 'Some text\n\n', sorting will add all those extra \n at the start, so we remove them here.
             stripped_text_from_html = stripped_text_from_html.replace("\n\n", "\n")
             stripped_text_from_html = '\n'.join(sorted(stripped_text_from_html.splitlines(), key=lambda x: x.lower()))
+### OPENAI?
+
+
+        # And here we run LLM integration based on the content we received
+        LLM_keys =  self.datastore.data['settings']['application']['ai'].get('API_keys', {})
+        if watch.get('LLM_prompt') and stripped_text_from_html and LLM_keys:
+            response = ""
+            try:
+                integrator = LLM_integrate(api_keys=LLM_keys)
+                response = integrator.run(
+                    provider="openai",
+                    model="gpt-4.1", #gpt-4-turbo
+                    message=f"{watch.get('LLM_prompt')}\n----------- Content follows-----------\n\n{stripped_text_from_html}"
+                )
+            except Exception as e:
+                logger.critical(f"Error running LLM integration {str(e)} (type etc)")
+                raise(e)
+                x = 1
+                # todo is there something special when tokens are used up etc?
+            else:
+                stripped_text_from_html = response
+               # logger.trace("LLM done")
+            finally:
+                logger.debug("LLM request done (type etc)")
 
 ### CALCULATE MD5
         # If there's text to ignore

changedetectionio/static/js/watch-settings.js

@@ -21,6 +21,7 @@ function request_textpreview_update() {
         namespace: 'watchEdit'
     }).done(function (data) {
         console.debug(data['duration'])
+        $('#error-text').text(data['duration']);
         $('#filters-and-triggers #text-preview-before-inner').text(data['before_filter']);
         $('#filters-and-triggers #text-preview-inner')
             .text(data['after_filter'])
@@ -37,9 +38,8 @@ function request_textpreview_update() {
     }).fail(function (error) {
         if (error.statusText === 'abort') {
             console.log('Request was aborted due to a new request being fired.');
-        } else {
-            $('#filters-and-triggers #text-preview-inner').text('There was an error communicating with the server.');
         }
+        $('#error-text').text(error.responseJSON['error']);
     })
 }
 

changedetectionio/static/styles/scss/parts/_edit.scss

@@ -1,3 +1,5 @@
+@use "_llm-prompt";
+
 ul#conditions_match_logic {
     list-style: none;
   input, label, li {