Biscovery RAG: an internal workspace for project knowledge

1The problem

• Past proposals, PDDs, and delivery files sit in drives and inboxes, so the next opportunity often starts from scratch.
• Searching by file name misses what the documents actually say.
• A general chatbot can invent projects that were never delivered.
• The library should open only to people the team has allowed in.

2What we did

1. 1Sign-in goes through Cloudflare Access: an email, then a one-time code, before the workspace opens.
2. 2The dashboard shows indexed projects and documents, knowledge chunks stored in PostgreSQL, searches run today, and whether the AI model is live — with shortcuts to search, upload, and the project library.
3. 3Upload one file or a batch: PDF, DOCX, Markdown, or text, up to 100 files and 100 MB each. Each file carries a project id, name, client, industry, document type, status, and technology tags.
4. 4Indexing runs in the background: queue, parse and segment, AI tags and embeddings, then metadata and vectors saved in PostgreSQL (pgvector).
5. 5The agent answers in English or Chinese from the project library. Thinking mode uses a slower model when the question needs more care. People can attach office files, archives up to 100 MB, or images up to 10 MB. Answers stay inside indexed projects; disk paths and internal playbook names stay out of the chat. Admins set how many rounds the agent may take, and which playbooks it may use.
6. 6Administrators allow specific login emails, assign a normal user or admin role, and can disable an account.

3Inside the Platform