AI · Knowledge
Knowledge & RAG on VegaDūta
Retrieval-augmented generation (RAG) on VegaDūta means your agents answer from your documents instead of guessing. You upload PDFs, DOCX files, HTML, images, or audio and video; the pipeline extracts the text (with OCR where needed), chunks it, embeds it, and serves it back to agents through a knowledge-search tool at answer time.
Embeddings are computed locally on CPU with sentence-transformers — your documents are not sent to a hosted embeddings API — and a cross-encoder reranker sharpens retrieval before the model sees it.
From file to searchable knowledge
The ingestion pipeline handles the formats businesses actually have, including scanned documents and recordings. Each stage is purpose-built rather than a single generic parser.
- Extraction for PDF, DOCX, HTML, and images
- OCR via Tesseract locally, with Google Vision as a cloud option
- Audio and video transcription via Whisper
- Chunking tuned for retrieval, then embedding and indexing
Embeddings stay on the platform
Vector embeddings are computed locally on CPU using sentence-transformers models. That is a deliberate privacy stance: document content is not shipped to a third-party embeddings API just to make it searchable. Combined with schema-per-tenant Postgres isolation, one tenant's knowledge never mixes with another's.
Retrieval that agents can use — and write to
Knowledge search is an agent tool, so retrieval happens inside the agent's reasoning loop rather than as a bolted-on lookup. A cross-encoder reranks candidate chunks so the model gets the most relevant passages, and agents can also write to the knowledge base, letting them capture new information as they work.
What is honestly not wired
AWS Textract and Azure Document Intelligence are not wired as extraction backends today; documents that would use them fall back to Tesseract OCR. If your documents are clean digital PDFs this rarely matters, but heavily degraded scans get Tesseract-grade OCR, not a cloud document-AI service.
Frequently asked questions
What file types can I upload to the knowledge base?
PDF, DOCX, HTML, and images, plus audio and video which are transcribed with Whisper. Scanned documents and images go through OCR — Tesseract locally by default, with Google Vision available as a cloud option.
Are my documents sent to an external AI service?
Not for embedding. Vectors are computed locally on CPU with sentence-transformers, so document content is not sent to a hosted embeddings API. OCR uses Tesseract locally unless you opt into the Google Vision cloud option, and tenant data is isolated schema-per-tenant in Postgres.
How do agents use the knowledge base?
Through tools: agents call knowledge search during a conversation to retrieve relevant chunks, which a cross-encoder reranks before the model answers. Agents can also write back to the knowledge base, so what they learn in one conversation can inform the next.
Does VegaDūta support AWS Textract or Azure Document Intelligence?
Not currently — those extraction backends are not wired, and such documents fall back to Tesseract OCR. Google Vision is available as the cloud OCR option today.
See it working in two minutes
The sandbox provisions a real tenant — describe an agent in one sentence and test it, no account, no card. Or browse ~90 industry workflow recipes to see what teams build.
