CrossMeet / Guide

Knowledge base and answer retrieval

Organize business knowledge as categorized questions and answers, retrieve it during meetings, and manage embedding engines and data paths.

Updated

The knowledge base (RAG) stores categorized questions and answers and retrieves relevant entries from conversation context. It is a Pro / Team feature, separate from document translation.

1. Create categories and entries

Open Knowledge Base from the app sidebar. Create categories such as “Product details” or “Delivery process.”

Each entry contains:

Field What to include
Question How someone might ask about the topic
Answer Verified information you can reuse
Keywords Names or terms that help identify the entry
Enabled state Whether the entry participates in retrieval

For example, ask “What needs confirming before delivery?” and record your actual checklist. This illustrates how to structure an entry; it is not a CrossMeet service commitment.

The current knowledge base manages Q&A entries and compatible JSON imports and exports. It does not provide a general upload flow that automatically turns arbitrary PDFs, Word files or slides into a knowledge base.

2. Configure embeddings

Embeddings turn text into vectors for similarity retrieval. Choose a compatible knowledge-base configuration in engine settings and allow entry vectors to finish generating.

  • For local embeddings, install the model and check hardware readiness.
  • For provider APIs, ensure configuration, connectivity and quota are valid.
  • Changing the embedding model can require regenerating existing vectors. Review entry status.

Vectors from different models are not automatically interchangeable.

3. Test retrieval

Use the knowledge base's search test with a representative question. Review the returned entries and use the available category, threshold and result-count controls.

If results are poor, check that entries are enabled and vectors are valid. Make questions and answers clear, remove outdated information and avoid contradictory entries. An optional reranking engine can refine result ordering.

4. Use it in meetings

Open knowledge retrieval from the meeting toolbar to look up answers using recent conversation context. Insert a selected answer into the reply field, then review and edit it before sending.

When automatic queries are enabled, finalized conversation segments can trigger background searches and show relevant-hit indicators. Retrieval is an aid; it does not mean every translation automatically receives your complete reference collection.

5. Where data goes

Entries are stored locally, but processing depends on configuration:

  • Cloud embeddings receive the entry and query text needed to generate vectors.
  • A cloud language model used to extract search intent receives relevant conversation context.
  • Cloud reranking receives the query and candidate text needed for ranking.
  • Translating an answer inserted into the reply field sends that text to the selected translation engine.

A fully local workflow requires every participating stage to run locally. Providers bill applicable usage separately. See privacy.

6. Import, export and maintain

Use JSON export to back up categories and entries, with or without vectors. After importing, check whether vectors match the selected model or need regeneration.

Back up before an overwrite import and review the interface's scope. Export files can contain business knowledge; handle them accordingly. Automatic cloud-drive, SharePoint or team knowledge-base distribution is not promised.

Glossary or knowledge base?

A glossary is for names, preferred translations and guidance. The knowledge base is for fuller answers and business explanations. Both require human review for important output.

Your next conversation

Start with a clearer understanding.

Explore Free, then add the tools your work needs.