Specifically: - Multi-channel deployment (Slack/Discord) is working but needs more testing - Some integrations (Teams, advanced webhooks) are in progress - Documentation is being expanded
The core functionality (RAG knowledge base, tool use, web deployment, self-hosting) is solid and ready to use.
Would rather launch early and iterate with community feedback than wait for "perfect"! Happy to be transparent about what's ready and what's coming.
I was focused on features and neglected the developer experience. That's a big miss.
I'm starting a new job tomorrow and won't have bandwidth to fix these immediately, but I'd love help from the community if anyone wants to contribute: - Unified docker-compose setup - Better documentation - Local model support (Ollama/llama.cpp)
All PRs welcome! Thanks for trying it out despite the rough edges.
When will you support AWS Bedrock?
Monetization: None planned. This is MIT licensed and I'm starting a new job tomorrow, so I won't have time to develop it commercially.
Why trust it exists in 6 months? You probably shouldn't rely on me maintaining it actively. That's why it's MIT licensed - anyone can fork and maintain it.
AWS Bedrock: Not currently supported. Would love a PR if someone wants to add it!
I'm sharing this more as "here's what I built while learning, maybe it helps someone" rather than "here's a product I'm committed to maintaining."
Yes, a lot of this was AI-assisted coding (Claude/Cursor), and I didn't clean up all the patterns. The exception handling is inconsistent, there are useless comments, and the code quality varies.
I'm the first to admit the codebase needs a lot of work. I was learning and experimenting, and it shows.
If you (or anyone) wants to improve it, I'd welcome PRs! The architecture/approach might be useful even if the implementation is rough.
Thanks for looking at the code and giving honest feedback - this is exactly the kind of thing I needed to hear.
If you already have a "large" product that's just very unfinished, that's not a MVP, you're asking people to be alpha testers. You either need to get it into a better state and make sure the docs and onboarding are exemplary, or break the product out into parts and ship them separately.
I got excited about the architecture and kept adding features. Classic scope creep. The feedback about breaking it into parts is really valuable - that would have been a smarter approach.
Starting a new job tomorrow, so I can't polish it further right now. But if anyone wants to fork it or collaborate on making it more production-ready, I'd be happy to help review PRs when I can.
Thanks for the honest feedback - exactly what I needed to hear.
HONEST CONTEXT:
This started as a personal learning project 6 months ago. I kept adding features as I learned, and it grew into something I thought might be useful to others. It's far from polished or complete.
I'm the first to admit: • The codebase has issues (yes, lots of AI-assisted code that needs cleanup) • Documentation is incomplete (some links are broken, sorry!) • Security features aren't fully implemented yet • Docker setup needs work (missing unified docker-compose) • Many features are still in testing/progress
I'm starting a new job tomorrow and won't have time to actively develop this for the foreseeable future. Rather than let it sit on my laptop, I wanted to share it with the community in case: 1. Others find it useful as-is for learning/experimentation 2. People want to collaborate and improve it 3. The architecture/approach helps someone building something similar
This is NOT a product I'm trying to sell or a startup. It's MIT licensed specifically so anyone can fork it, improve it, or learn from it.
LOOKING FOR COLLABORATORS:
If anyone wants to actively work on this project, I'd love that! The codebase needs: - Security hardening - Better documentation - Cleaner exception handling - Unified docker-compose setup - Local model support (Ollama/llama.cpp) - Better testing
I'm happy to answer questions about the architecture and will try to review PRs when I can, but I can't commit to active development right now.
Thanks for the honest feedback - it's exactly what I needed to hear. If this helps even one person learn something or sparks ideas for better implementations, that's a win.
achushankar•3mo ago
The Problem: I kept needing AI that could both answer questions from company knowledge AND take actions (send Slack messages, trigger workflows, call APIs). Existing solutions either don't support agentic tool use well, lock you into their cloud, or require weeks to build from scratch.
What Syllabi Does (Three Pillars):
1. INTEGRATE ANY TOOLS • Call webhooks & custom APIs • Send Slack messages, emails, calendar events • Trigger workflows in external services • Connect YOUR custom tools via API/webhooks • AI intelligently decides WHEN and HOW to use each tool (that's the agentic part)
2. KNOWLEDGE BASE (RAG) • Transform docs, videos, websites into knowledge base • PDFs, Google Drive, Notion, Confluence • Advanced RAG with source citations • Click citations to see exact passages highlighted in original documents • Multi-format processing with smart chunking
3. DEPLOY ANYWHERE • Embed widget on any website • Slack & Discord bots • Microsoft Teams (coming soon) • Standalone web app • REST API for custom integrations • One agent, multiple channels
Key Technical Features: - MIT licensed, self-hosted, privacy-first - Modern AI models (latest GPT-4, GPT-4o, o1 series) - More providers coming (Anthropic, local models) - Agentic tool selection with function calling - Channel-agnostic core with adapter pattern - Async job queue for document processing - Plugin system for custom skills
Tech Stack: Next.js (frontend), Python FastAPI (backend), PostgreSQL, Supabase, OpenAI API
Use Cases: • AI course assistant that answers questions AND books office hours • Support bot trained on docs that can create tickets in Linear/Jira • Team knowledge base in Slack that triggers workflows • API docs helper that generates AND runs code examples
Architecture Highlights: - Modular design with clean separation of concerns - Row-level security for multi-tenancy - Docker deployment for easy self-hosting - Comprehensive API for custom integrations
I started building this 6 months ago because every project seemed to need the same thing: an AI that could access knowledge AND take actions, without vendor lock-in or per-message pricing.
Website: https://www.syllabi-ai.com/ GitHub: https://github.com/Achu-shankar/Syllabi Docs: https://www.syllabi-ai.com/docs
Would love feedback from the HN community – especially on: - Agentic AI architecture approaches - Tool use and function calling strategies - Multi-channel deployment patterns - Self-hosting and security best practices
Happy to answer questions about the tech stack, RAG implementation, agentic tool selection, or anything else!
diffeomorphism•3mo ago
What kind of security guarantees do you have?
It seems to meet that your "problem" usually is unanswered on purpose:
https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/
This has access to sensitive knowledge, tool use and exfiltration. So, the tech seems nice, but I doubt I could ever get permission to deploy this.
achushankar•3mo ago
Is a serious concern, especially in enterprise environments.
Currently, this has: - Row-level security in Supabase - API key auth - Rate limiting
But it does NOT have: - Comprehensive audit logging - Fine-grained permission controls - Tool execution sandboxing - Data loss prevention
You're right that this shouldn't be deployed in production with sensitive data without significant security hardening. I should have been clearer about that.
Thanks for the link to Simon's article - very relevant. This is more suited for learning/experimentation than production use with sensitive data right now.
If anyone wants to work on security features, I'd be happy to collaborate on that!