20250417

Notes

Newly added Retrieval-Augmented Generation (RAG) pipeline which enhances response quality by retrieving relevant information from multiple sources before generating answers. The pipeline sources its information from Wikipedia articles, web search results, and news sources.

Implements a lightweight but effective vector embedding system for semantic similarity matching. Uses a hybrid approach combining keyword matching and vector similarity for more robust retrieval performance. Includes recency detection that prioritizes fresh information for time-sensitive queries. The pipeline typically adds about 300 tokens of highly relevant context to user queries.

├── managers/search/
│   ├── service.ts       # Orchestrates content retrieval from multiple sources and integrates with vector search
│   ├── vectorstore.ts   # In-memory vector store for document embedding, chunking, semantic search, and recency boosting

Users can enable web search with + search inside a thread. Web search can be disabled with + model.

20241215

Notes

Initial release featuring LiteLLM integration that enables support for 40+ model providers. Support for OpenAI-compatible APIs enables connections to a range of open-source inference engines and wrappers.

The LiteLLM proxy is bundled as part of the docker-compose.yaml.

Reorganized the code to make the project structure simpler:

commands/
├── base.ts          # Command interfaces and base classes
├── handlers.ts      # Unified implementation of all commands and services
└── registry.ts      # Centralized command registration and routing
managers/generation/
├── core.ts          # Message orchestration and routing
├── formatter.ts     # Output formatting and pagination
└── generator.ts     # Response generation and streaming
utilities/
├── error-handler.ts
├── index.ts         # Combined types and exports
└── settings.ts

Cleaned up the file structure to make future updates simpler and smoother. The new organization keeps everything in logical places while making the code easier to work with.

Fixes

Added handling to address an error that would cause the bot to continuously check the cache for a message that cannot be found:

Error updating message: Error [ChannelNotCached]: Could not find the channel where this message came from in the cache!

Fixed an issue where messages would be copied into different conversation histories when + was used to switch models mid-conversation.