SovrGPT Docs

Knowledge base (RAG)

RAG knowledge base in SovrGPT: upload your own documents and get answers with sources, strictly separated per organisation.

The knowledge base is your organisation's searchable document collection. Technically it is RAG (retrieval-augmented generation): before answering, SovrGPT searches your documents, hands matching passages to the model and cites them as sources. That keeps it verifiable where a statement comes from, for example in public administration, legal and compliance. No organisation sees another's knowledge.

Choosing knowledge sources per chat

You choose the source in the input area with the button that shows the current mode, for example "Wissensbasis":

ModeBehaviour
OffNo knowledge base, a normal chat.
Knowledge baseEnrich answers with your knowledge base.
Knowledge base onlyAnswer exclusively from the knowledge base, without world knowledge.
WebWeb search only, without the knowledge base.
Knowledge base + webCombine the knowledge base and web search.

The organisation sets a default mode, and every chat can change it. The web modes use Brave Search, a provider based in the USA. If such a mode is the default, web search is on in every chat without its own setting.

Managing documents

Under Einstellungen → Wissen (RAG) (settings → knowledge) you upload files (PDF, Word, Excel, text, Markdown), see the document list, delete individual documents and keep an eye on the quota. On upload, SovrGPT extracts the text, splits it into sections and stores them as vectors in our database in Frankfurt.

Every member of the organisation can upload documents. Members delete the documents they uploaded themselves; owners and admins can delete any document.

How the search works

  1. Hybrid search: keyword search (Postgres full text) and semantic search (pgvector) are merged.
  2. Reranking: a language model reads the question and up to 20 of the best hits and determines which passage answers the question best; the order follows from that. In a measurement on our own documentation this put the right passage in first place considerably more often than semantic search alone.
  3. Citing: the top passages go to the model. The answer marks them with [#1], [#2] …, and below the answer each cited marker shows the document and the section.

If the search cannot run for technical reasons, SovrGPT says so openly and does not claim that your documents contain nothing relevant. In Knowledge base only mode you then get a request to try again rather than an answer from world knowledge. In the other modes the answer states that it was written without your sources. If only reranking fails or takes longer than three seconds, the hits arrive in the order of the hybrid search.

Embedding (documents on upload, questions on search) runs at STACKIT in Germany, with no wait for a server to start.

Who reranks the hits

OrderModelData centre
1.Gemma 4 31B · STACKIT (stackit-gemma-4-31b)Germany
2., only when Gemma 4 31B · STACKIT is busy, faulty or unreachable, or is hidden in your organisationQwen 3.8 27B · Irland/Finnland (tensorx-qwen3.8-27b) in the European partner data centreIreland, Finland
Fallbackorder of the hybrid search–
  • Only approved models: your passages only go to models that are visible in your organisation. A hidden model is skipped; if both are hidden, the order of the hybrid search applies. To keep passages out of the partner data centre, hide “Qwen 3.8 27B · Irland/Finnland” – not “Qwen 3.8 27B”, which runs at STACKIT.
  • Three seconds for everything: if none of the models has answered by then, the order of the hybrid search applies. The answer does not wait any longer.
  • Only your own passages: only hits from your organisation are reranked. Reranking only changes the order; it never adds passages.
  • Nothing in the logs: when reranking, SovrGPT logs only numbers and identifiers, neither your question nor any passages. By contract, the European partner data centre does not store inputs.
  • Cost: reranking is charged at the price of the model that answered – roughly 10,000 to 11,000 input tokens and one output token per search. If none answers, nothing is charged. Embedding is charged too: your question on every search and the passages when uploading, at the embedder's price (prices: GET /v1/pricing). Your spending limit applies to both.

Through the API, POST /v1/rerank still reranks with bge-reranker-base, self-operated (see Models); the knowledge base in the chat no longer uses this model.

Separation between organisations

Every row carries the org_id of its organisation. Row-level security lets only members read; writing happens exclusively on the server, never directly from the browser.

Availability (plans)

FeaturePlan
Knowledge basePro and above (up to 200 documents, 500 MB); enterprise by agreement
Web search as a sourceall plans

Knowledge base vs. file attachment

The knowledge base is permanent and searchable in all of the organisation's chats. A file attachment applies only to that one conversation.

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

Knowledge base (RAG)