SovrGPT Docs

Knowledge base (RAG)

Upload your own documents and have answers backed by them — tenant-isolated retrieval-augmented generation on EU infrastructure.

The knowledge base is an organisation's searchable document collection. Technically it is RAG (retrieval-augmented generation): before answering, SovrGPT searches your own documents, hands the matching passages to the model and backs the answer with sources. The model does not guess — it looks things up, and says where it looked.

What it can do

  • Your own documents as a knowledge source: PDF, Word, Excel, text, Markdown.
  • Sourced answers: passages are cited — traceable, which matters for public administration, legal and compliance.
  • Controllable per chat: you decide per conversation whether the knowledge base is used, only the knowledge base, only the web, or a combination.
  • Strictly tenant-separated: one organisation never sees another's knowledge.

Choosing knowledge sources per chat

You choose the knowledge source in the composer (the RAG: pill). There are five modes:

ModeBehaviour
OffNo knowledge base — a normal chat.
Knowledge baseEnrich answers with your knowledge base.
Knowledge base onlyAnswer exclusively from the knowledge base (no world knowledge).
WebWeb search only, without the knowledge base.
Knowledge base + webCombine the knowledge base and web search.

The organisation sets a default mode; every chat can override it. "Knowledge base only" is ideal when answers must rest exclusively on approved in-house documents.

Managing documents

Under Settings → Knowledge (RAG):

  1. Upload files (PDF/Word/Excel/text/Markdown). The text is extracted, split into sections, turned into vectors and stored.
  2. Review the document list, delete individual documents.
  3. Keep an eye on the quota (number of documents and total size).

Content is stored encrypted in the EU and stays confined to that organisation.

How the search works

SovrGPT combines two methods and then re-orders the hits:

  1. Hybrid search — keyword search (Postgres full text) and semantic vector search (pgvector) are merged.
  2. Reranking — a cross-encoder (bge-reranker-base) sorts the best hits by actual relevance.
  3. Injection + citation — the top passages go to the model, and the answer points back to where they came from.

Embedding and reranking run on EU-sovereign endpoints — no hyperscaler, no data outside the EU.

Tenant isolation (non-negotiable)

Strict separation per organisation is the core requirement of this feature:

  • org_id on every RAG table — no record without an organisation.
  • Row-level security: reading only for members of the organisation; writing exclusively server-side (never directly from the client).
  • Encrypted storage in EU data centres.

Availability (plans)

FeatureFrom plan
Upload to and query the knowledge base (RAG)Pro
Web search as a sourceall plans

If your tier is not sufficient, the RAG: pill shows an upgrade hint.

Knowledge base vs. file attachment

  • Knowledge base = the organisation's persistent document collection, searchable across all chats.
  • File attachment in a chat = transient: a file is only taken into account for the ongoing conversation and is not indexed permanently.

Related topics: Marketplace · Connecting your own systems · GDPR & compliance.

Knowledge base (RAG)