A self-hosted research notebook — upload your sources, ask questions across them, get cited answers. Like NotebookLM, but yours.
AI is powerful but generic. It doesn't know your documents, your research, or your context.
Google's NotebookLM solves this — upload PDFs, docs, notes, and URLs, then ask questions that pull from your sources with inline citations. The result is research that's grounded, verifiable, and specific to you.
But NotebookLM is Google's product. Your data lives on their servers. The interface is their design. The feature set is their roadmap. And there's no API to connect it to Hermes or any other agent.
The insight: The core pattern — ingest sources, generate embeddings, answer with citations — isn't Google-exclusive. It's RAG (retrieval-augmented generation) with a good UI. And we can build it, own it, and connect it directly to Hermes.
A self-hosted research notebook powered by your own documents.
Standard RAG stack, owned and operated by IT Pro Partner.
| Layer | Technology | Notes |
|---|---|---|
| Ingestion | Unstructured + Marker | PDF, DOCX, TXT, HTML parsing with OCR fallback |
| Chunking & Embedding | ChromaDB or Qdrant | Semantic chunking, multiple embedding models (local or API) |
| Query + RAG | LangChain or custom FastAPI | Hybrid search (semantic + keyword), reranking |
| LLM | LiteLLM proxy | Routes through our existing admin-ai provider pool |
| Frontend | SvelteKit or FastAPI + HTMX | Notebook-style UI: source panel, chat panel, citation panel |
| MCP Server | FastMCP (Python) | Exposes query tool to Hermes and other agents |
| Auth | Central Auth (auth2) | SSO with existing ITPP identity, multi-tenant ready |
Three phases, starting with the Hermes integration.
Phase 1 — Hermes Skill (1 week): Build a `research-notebook` Hermes skill. Ingests files from disk, chunks and embeds them locally (ChromaDB), exposes query tool. No UI — everything through Hermes conversation. Proves the RAG pipeline works.
Phase 2 — Web UI (2-3 weeks): Standalone SvelteKit app. Notebook-style interface with source sidebar, chat panel, and citation highlights. Multi-notebook support. Deploy on app3 behind auth2.
Phase 3 — MCP Server + Multi-tenant (2 weeks): Expose as MCP server so any agent (Hermes, Claude Code, Codex) can query vaults. Add tenant isolation. Pricing tiers for external users.
This isn't just an internal tool. It's a product.
| Tier | Target | Value Prop |
|---|---|---|
| Internal | ITPP operations | All ITPP docs, proposals, client research in one queryable vault. Hermes answers every question with citations to our own knowledge base. |
| Client add-on | Existing MSP clients | Per-client vault with their documentation. "Ask your IT manual anything." $50-100/mo per client. |
| SaaS standalone | Researchers, law firms, consultants | Self-serve SaaS. Upload sources, query, export. $20-100/mo based on source count and queries. Compete with NotebookLM on privacy + API access. |
Why this works: NotebookLM proved the demand. People want AI that works with their documents. Google's version is free, but you can't self-host it, can't connect an API, and your data trains their models. Privacy-conscious professionals will pay for a self-hosted alternative with API access.
Bottom line: Build the Hermes skill first (Phase 1, ~1 week) to prove the RAG pipeline. If it's useful internally, that alone justifies the build. Everything beyond that is product revenue.