openwebui-filters

Filters for Open WebUI that handle context management and response formatting.

Filters

Token Saver (filters/token_saver.py)

Cleans up the message history before it reaches the LLM.

• Timestamp injection: Fetches real timestamps from the Open WebUI API and prepends them to historical messages; injects the current date and time into the latest message.

• History trimming: Keeps only the last N messages (default: 50) to reduce token usage.

• Image pruning: Strips image attachments from older messages so text-only models (e.g., DeepSeek R1) do not process multimodal chat history unnecessarily.

• Token optimization: Disabled by default. Removes filler words (e.g., "just", "really", "very") from user messages while preserving code blocks, Markdown formatting, and quotes. This reduces token counts without sacrificing critical information.

Valve	Default	Description
priority	0	Execution order (higher equals later execution).
n_last_messages	50	Messages to keep (user and assistant each count as 1).
ignore_images_after_n_messages	10	Strip images from messages older than this (0 strips all).
api_base_url	http://127.0.0.1:8080	Open WebUI API base URL.
inject_system_instructions	True	Adds a system prompt explaining the timestamp format to the LLM.
enable_token_optimizer	False	Toggles filler-word removal.
min_words_to_optimize	60	Skips optimization below this word count.
max_words_to_optimize	2000	Skips optimization above this word count.
optimize_current_only	True	Only optimizes the latest message.

Requires httpx>=0.24.0 (included in the default Open WebUI Docker image).

Concise Output (filters/concise_output.py)

Injects system-level instructions that push the LLM toward concise, structured output.

• Sets a soft target word count (default: 240).
• Enforces bold-caps headers and bullet lists.
• Suppresses preambles (e.g., "As an AI") and filler phrases (e.g., "Certainly!").

Valve	Default	Description
priority	0	Execution order (higher equals later execution).
target_word_count	240	Soft word-count target; the model can exceed it when accuracy requires.

No extra dependencies are required.

Installation

Requirements: Open WebUI v0.8.4 or higher.

Import from URL

1. Navigate to Admin Panel → Functions → Import From Link.
2. Paste the preferred raw URL:
   • https://raw.githubusercontent.com/fractuscontext/openwebui-filters/main/filters/token_saver.py
   • https://raw.githubusercontent.com/fractuscontext/openwebui-filters/main/filters/concise_output.py
3. Enable the filter.

Manual Installation

Copy and paste the code from the respective filter files directly into the Open WebUI Admin Panel.

License

MIT License