open-webui-filters/CHANGELOG.md

OpenStreetMap Tool

2.1.0

• New feature: navigation. LLMs can provide navigation directions and answer questions about distance between two places. Works best with Qwen and Mistral. Llama3 seems to require a very specific request to use the navigation tool.
• New POI category: tourist attractions. Uses new a new ranking system to try and surface more prominent attractions.
• Fall back to Haversine distance if OpenRouteService cannot calculate a navigation distance.
• Cache ORS route information.

1.3.1:

• Handle bad or unclear addresses by having the model tell you.
• Fix a swallowed exception that resulted in no results, but no notification of finding no results.

1.3.0:

• Emit one citation per result, with nicely formatted text.

1.2.0:

• Update citations for compatibility with 0.4.3.
• Very basic prettification of citation by converting the Markdown to HTML.
• Show friendlier name for citation when searching for POIs (usually address instead of GPS coordinates).

1.1.3:

• Better friendly display names when an address is available.

1.1.2:

• Fix misleading Nominatim resolution complete event that was communicating as still resolving.

1.1.1:

• Send resolution complete event when pulling Nominatim info from cache.
• Send error event when headers not set for Nominatim search.

1.1.0:

• Fix bad check in fallback name assembly behavior.
• Slightly alter functions to improve specific place search.

1.0.0:

• Breaking change: Nominatim URL must now be set to the root domain. The model will warn you about this.
• Caching: Almost all data that the tool fetches is now cached in a JSON file (/tmp/osm.json). This reduces load on public OSM services.
  • Future updates will allow more control over cache behavior.
  • The only thing that is not cached at all is the Overpass Turbo search, that actually finds POIs.
• Handling of unnamed points of interest: many smaller POIs like neighborhood playgrounds do not have names. The tool will now look up addresses for theses POIs using Nominatim.
• Citations: The tool will now send citations when the indicators setting is turned on. The citation contains the results of the search given to the LLM.

0.9.0:

• Integrate OpenRouteService to allow calculation of more accurate distances based on the distance of travel rather than "as the crow flies" distance.
• If OpenRouteService is not enabled, the original method will be used for calculating distance.
• ORS is not enabled by default.
• Properly extract amenity type of leisure locations (playgrounds, parks, etc) in search results.

0.8.0:

• Added ability to find specific stores and businesses near coordinates. Helps LLM answer questions about arbitrary businesses that it might not find with the pre-defined search functions.
• Pull shop type from Nominatim results when class=shop. Improves accuracy of searching for specific places by telling LLM the type of shop.

0.7.0:

• Emit event to web UI, showing what the tool is doing.
• This can be toggled, as a valve setting (not a user valve).

0.6.3:

• Add tool function for resolving GPS coordinates to address.
• Improve handling of questions like "Where is X near me?"
• Add basic logging (will be toggleable in a future version).

0.6.2:

• Override doityourself store to hardware.

0.6.1:

• Small adjustment to single location result instructions.

0.6.0:

• Dramatically improved single location lookup (e.g. asking "where is Landmark/City?")
• Add a catch-all handler that the model can pick if the user wants to searh for something not added to the tool. The function will force the LLM to tell the user what the OSM tool can find.
• Added more findable POIs: Hardware stores/home improvement centers, electrical and lighting stores, consumer electronics stores, and healthcare (doctors' offices, hospitals, health stores).

0.5.1:

• Remove instruction to prioritize nodes over ways in search results.

0.5.0:

• Support Way results. This makes searching much more accurate and useful. Many map features are marked as "ways" (shapes on the map) rather than specific points.
• Drop support for "secondary results," and instead return Ways that have enough information to be useful.

0.4.0:

• Complete rewrite of search result handling to prevent incorrect OSM map links being generated, and bad info given.
• New setting: Instruction Oriented Interpretation. Controls the result interpretation instructions given to the LLM. Should help certain models (Hermes Llama 3.1) give more consistent results.
• The Instruction Oriented Interpretation setting is also a user valve that can be controlled on a per-chat basis.
• Added ability to search for: public transit, schools/universities, libraries, bike rental locations, car rental locations.

0.3.0:

• Add search for alcohol and cannabis/smart shops.
• Increase hotel search radius to 10km.

0.2.5:

• Better accuracy in turn by turn conversations: model encouraged to specify the name of the city and country when searching, so it's less likely to report results from a different country.

0.2.4:

• Actually make use of the limit parameter when searching Nominatim.

0.2.3:

• Allow leisure tags in nominatim fallback results.

0.2.1:

• Only print secondary results if they exist.
• Remove extraneous print logging.

0.2.0:

• Include Ways as secondary information for improved search results.
• Added searching for swimming and recreational areas.

0.1.0:

• Initial release.

Checkpoint Summarization Filter

0.2.2:

• Fix imports for OpenWebUI refactoring.

0.2.1:

• Fix None - int error when setting num_ctx in a chat, and then unsetting it.

0.2.0:

• Update for newer versions of OpenWebUI (0.3.29+).

0.1.0:

• Initial release.

Narrative Memory Filter

0.0.2:

• More reliable way of getting chat IDs.

0.0.1:

• Initial proof-of-concept release.

GPU Scaling Filter

0.2.2:

• Fix imports for OpenWebUI refactoring.

0.2.1:

• Fixes for internal OpenWebUI refactoring.

0.2.0:

• Fix filter not working when using base models directly.
• Do not fire the filter if model is not owned by ollama.

0.1.0:

• Initial release.

Output Sanitization Filter

0.2.0:

• Prevent errors when an LLM reply is updated by the filter (i.e. return body from the outlet function).

0.1.0:

• Initial Release

Collapsible Thought Filter

0.2.0:

• Fix issue with output disappearing.

0.1.0:

• Initial release.

Artificium Thinking Filter

0.1.1:

• More reliable discovery of thinking and output sections.
• Remove dead code.

0.1.0:

• Initial release.

Gemini Tool

0.1.1:

• Do not correct URLs when following redirects.
• Improve result handling.

0.1.0:

• Handle redirects.

0.0.1:

• Initial release with basic protocol handling.