¶ Minime-T5 / Text2Vec
Minime-T5 is a lightweight local AI service using Google's Flan-T5-Small model.It is fast and has low memory usage. It can be run on CPU or GPU:
- CPU Mode: ~400 MB RAM
- GPU Mode: ~400 MB VRAM (CUDA)
You can install it within the CHIM.exe launcher under Install Components. The service runs locally on port 8082 and provides several endpoints for different tasks.
¶ What Minime-T5 Does
1. Action Triggering (/command)
Every user input is analyzed to detect if a specific action should be triggered. Currently detects:
- TakeASeat - "Can you sit down?"
- ReadQuestJournal - "What's my current quest?"
- ExchangeItems - "I want to give you something"
- Attack - Explicit attack orders
This is especially useful with less sophisticated LLMs that might miss action cues.
2. Memory Detection (/extract & /detectMemory)
Not every conversation needs to trigger a memory lookup. Minime-T5 analyzes text to determine:
- Whether a memory search should be triggered
- Extracts salient keywords for memory retrieval
You must enable AUTO_CREATE_SUMMARYS in the profile for this to work.
3. Goal/Task Extraction (/task)
Detects when a new goal or mission is being set. For example:
- "Our new goal is to retrieve all Words of Power"
- Minime extracts "retrieve all Words of Power" and adds it to the current mission table
This is the SetCurrentTask functionality that keeps AI NPCs aware of your objectives.
4. Topic Extraction (/topic & /posttopic)
Extracts the subject/topic from sentences for:
- Oghma Infinium - Determining what lore topics to inject
- Post-conversation analysis - Understanding what was discussed
5. Scene Analysis (/ambient)
(Experimental)
Analyzes dialogue history to detect scene mood/genre for dynamic responses.
¶ Oghma Infinium Integration
When both Minime-T5and Oghma Infinium are enabled:
- Player input is sent to Minime-T5 to extract the topic
- The topic is searched in the Oghma knowledge database
- Relevant lore is injected into the AI NPC's context
- NPCs can now speak knowledgeably about Elder Scrolls lore!
Alternative:You can use OGHMA_CUSTOM
instead of Minime-T5 for topic extraction, which uses your configured LLM instead (more accurate but uses tokens).
¶ Text2Vec (Embedding Service)
Text2Vec provides vector embeddings for semantic search in:
- Memory Retrieval - Finding relevant past memories
- Oghma Knowledge Search - Finding relevant lore entries
Two Modes:
- Local - Uses sentence-transformers/all-MiniLM-L6-v2 via a local TxtAI endpoint
- Remote - Uses OpenAI's embedding API
Configuration in Global Settings:
- MEMORY_EMBEDDING ENABLED - Toggle the embedding system
- TXTAI_URL - Endpoint for local Text2Vec service
- USE_TEXT2VEC - Enable vector-based retrieval
- MEMORY_CONTEXT_SIZE - How many memories to retrieve
- MEMORY_BIAS_A/B - Fine-tune relevance scoring
You can view the Text2Vec embeddings under Memory Summaries in the web interface.
