blob42/Instrukt v0.6.2

🚀 v0.6.2 - coding assistant with local embeddings from any code base

on GitHub

This release brings a major update to document indexes and embeddings
management 📚 .

Note that AI agents still rely on OpenAI for now. The next milestone will bring support local LLM models.

Index management UI:

• Using custom and local embedding functions.
• Implemented DirectoryLoader for scanning and indexing any source code base with preliminary support for language parsing (currently python, js) using LanguageParser from langchain.
• Indexing and searching multiple directories.
• Scanning and filtering content with custom glob.
• CUDA/CPU embeddings with SentenceTransformers.

Conversation UI:

• Many improvements to the UX giving more space to agent outputs.
• Fuzzy select and edit source documents used for retrieval Q&A.
  Selection is sent as stdin to the program exported in $SELECTOR
• Edit the agent reply or input using an external $EDITOR.

full changelog 0.6.2

Added

• output_parsers.parser_lib.get_rich_md to sanitize agent output to markdown
• wip: REPL input modes and auto completion
• wip: Programming assistant agent
• Progress bar for async events on tools
• select and edit source documents used for retrieval Q&A with ctrl+p
• Retrieval Q&A with agents returns source documents
• ProgressBar protocol and wrapper to use Textual progress bar in a thread safe way
  and hook into tqdm update events.
• Edit input using an external $EDITOR with ctrl+e
• Capture and redirect logging/output to Instrukt console widget.
• Generate sphinx doc for online and offline reading from within the app.
• wip: offline doc reader: jump to anchors (Textualize/textual#2941)
• Help screen for common keybindings with ?
• UI: reusable action bar widget.

index management:

• Embeddings: added bge-base embeddings option
• Scan and load multiple PDFs from a directory under a single collection
• AutoDirLoader: scan and load a directory, auto detects file types and assigns
  the appropriate splitter based on the detected content type.
• link (patch) the index console progress bar to tqdm updates
• progress bar for loading, splitting and indexing files
• added a local file system path selection UI
• chromadb: save/restore used embedding function. You can have multiple indexes using
  different embedding functions.
• Choose to use local embeddings when creating index.
• TODO: detect local embeddings when loading an index.

Changed

• Refactored the prompt input and console output for better UX.
• Bumped textual to version v0.34.0 (TextLog -> RichLog)
• Improved key bindings
• Memory mixin to handle retrieval answers with source documents
• Retrieval uses MMR search algorithm by default
• Improved form validation
• Upgraded dependencies
• ChromaDB: no more manual call to persist

Fixed

• index: async delete indexes: ensure deletion happens after index is loaded.
• improved iPython dev console: avoid term repaints until end of session.
• Fast preloading of messages when switching agent tab.
• Chroma: share a single client for all indexes
• Explicit dependency on sentence-transofmers library for local embeddings.
• Many fixes related to dependency updates