Demo

Sourced answers (RAG)

Answers grounded in your documents, with their sources

Hybrid search, contextual enrichment and verified citations: every answer points to the excerpts that support it.

Try sourced answersDiscuss your project
Which documents do I need to provide for a reimbursement request?

Attach the paid invoice and the signed request form 1. Above €500, prior approval from your manager is also required 2.

Sources

  1. 1Reimbursement procedure.pdfPDF“…every request must include the paid invoice and the signed form…”
  2. 2Internal memo no. 12.docxWord“…above €500, the manager’s prior approval is required…”
Interface diagram: sourced answer in the conversation portal (sample content).

The problem

Why it matters

On its own, a language model answers with confidence, even when it is wrong. In a professional setting, an answer must be verifiable.

Retrieval-Augmented Generation (RAG) means finding the relevant passages in your documents, then asking the model to answer from those passages and cite them.

In practice

  1. Understand the questionThe question is rephrased into several variants so that differently worded passages are not missed.
  2. SearchSemantic (vector) search and full-text search combined, then weighted by each knowledge base’s priority.
  3. WriteThe agent selects the useful excerpts, can dig deeper into a document, then writes its answer.
  4. CiteOnly excerpts actually retrieved can be cited; they are numbered [1], [2]… and shown with their source document.

Capabilities

What you can do

Each capability is shown with its actual availability.

  • Hybrid semantic + full-text searchRank fusion (RRF), on PostgreSQL and pgvector.Available
  • Contextual enrichment of excerptsEach excerpt is indexed with a sentence of context about its document.Available
  • Multi-query rephrasingAvailable
  • Knowledge base prioritiesFrom 1 to 10: in case of contradiction, the higher-priority source prevails.Available
  • Verified citationsDocument name, excerpt, original URL for web pages, viewing permissions per knowledge base.Available
  • “Insufficient evidence” warningShown when no relevant source was found.Available
  • Deep searchDedicated mode, with a specific model.Available
  • Model-based result rerankingAvailable when the internal model is OpenAI.Requires setup
  • Complementary web searchVia the Serper service (key required).Requires setup

Benefits

What you gain

Verifiable

Users can trace back to the cited excerpt and document in one click.

Fewer fabrications

The model answers from your content and says when it cannot find anything.

Source hierarchy

An official reference can take precedence over secondary documentation.

Requirements and limitations

Knowledge bases

Import your files, index websites, transcribe audio and video, and screen every document before it feeds into answers.

Language models

OpenAI, Mistral, Anthropic, Gemini or a model you host yourself: a choice per assistant, encrypted keys and a benchmark to decide based on facts.

Monitoring and improvement

Metrics, conversation review, AI analysis, reports, live takeover by an agent and a multichannel ticket center.

Next step

Ask the demo assistant questions and look at the citations.

Try sourced answersContact us