Retrieval-Augmented Generation

Live

Customers ask Initech's help bot about plans and policies the model was never trained on. Your call: how does it answer from Initech's own documents instead of guessing?

Key idea: RAG = search your own data first, then let the model answer from it.

Things to try

  1. Ask about Growth plan pricing with Retrieval off, then on. Without it the model has to guess or decline.
  2. Ask the SSO + EU hosting question with Top-k = 1. One chunk can't answer a two-part question.
  3. Set Top-k back to 3, turn on Reranking and compare the retrieved order in the panel with the reranked order.
  4. Turn Reranking off and ask something the docs don't cover ("Is there a Linux app?"). A grounded model should say it doesn't know.
Try a prompt
Enter to send · Shift+Enter for a new line

Change how it works, then send again

no retrievalindexed onceQuestionyour promptEmbedding modeltext → vectorpplx-embed-v1-0.6bVector store20 doc chunksKnowledge baseInitech CRM docsTop-k retrievalmost similarRerankeroffAugmented promptquestion + chunksLLMwrites the answerministral-8bAnswerwith citations
Send a message to watch it run

Retrieved chunks · by similarity

Ask a question about Initech CRM to see which doc chunks get pulled in.

Behind the scenes

Send a message and every step the system takes will show up here.