Home

Kollektiv Roadmap

Roadmap is subject to change.

🚂 v0.2.0 (WIP) - Open-source "Ask AI" for your content

Big idea

0.2.0 will give the user the easiest way to chat with web content (primarily documentation) with a nice, simple web UI (based on Chainlit). It will be offered in two options:

• open-source solution for DIY enthusiasts - free to use and setup on your own
• a showcase web app with a set of indexed documents focusing on AI agents, AI, LLM framework - covering the full lifecycle

Deployment

• Self-hosted: clone repo & setup your self
• Web-app: hosted solution with +30 AI / LLM / AI-agent focused libraries to work with

User flows

1. Index web content via @docs command

⚙️ Features

1. Automate ingestion pipeline

   • automate the pipeline from URL to embedding and loading
   • simplify db reset
   • index and do not reload documents that exist

2. Command support

   • @docs -> parse and load the parsed docs into db
   • @get -> get, parse and store a single URL

3. Chunking improvements:

   • Add chunk summary generated by LLM
   • Improve chunking quality (token, header preservation)
   • Try out Unstructured.io

4. Git sync

   • Allow syncing of a Git repo a-la Claude enterprise style
   • @repo -> setup indexing of a github repo

🔎 Search Relevancy

1. Improve stability of RAG calls

   • LLM should decide in a smart way when RAG is needed
   • Search re-tries with different query

2. Setup basic eval suite for RAG

   • Scope: a. Anthropic docs b. Supabase docs c. LlamaIndex docs
   • Evaluation metrics defined (e.g., accuracy, relevance, speed)
   • Automated test suite implemented
   • Baseline performance established for each doc set

🖥️ UX Improvements

1. Output streaming
2. Tool result streaming / output
3. Improved terminal output (using colorama)
4. Chainlit UI implementation

🏗️ Refactoring

1. Research the need to transition to LangChain [WON'T DO -> LCEL, WHY RE-INVENT THE WHEEL?]
2. Refactor code to use LlamaIndex for search & retrieval

📊 Research

1. Setup basic telemetry for open-source app with public dashboards

---

Future Releases (Backlog)

V0.3.0 - Chat with a GitHub repo via a web interface V0.4.0 - Confluence agent to chat with company docs V0.5.0 - Open-source plugin that can be installed on any website to instantly give it LLM + RAG over site content, securely, privately

Completed Releases

🚂 V0.1.0 - Initial release

Initial release of Kollektiv (called OmniClaude back then) with the following features:

• crawling of documentation with FireCrawl
• custom markdown chunking
• embedding and storage with ChromaDB
• custom retrieval with multi-query expansion and re-ranking
• chat with Sonnet 3.5 with rag search tool