Prompt Flow Overview

2 min read

Here is a detailed overview of the lifecycle of a prompt, using websearch:

Example Prompt:

I am Marie Dubois the IT director of Banque Genève, owned by the Audemars Family, with assets under management of 60 million CHF, and would like to know how my roles and responsibilities might evolve during the next 10 years. Please provide examples of IT managers in the same situation and show me the issues they are or will be facing in the future. Also, can you find real names of asset managers who are in the same situation as me?

Steps →

  1. Prompt

  1. LLM

  1. Web Search

  1. 2nd LLM Call

  1. Output Response

What happens?

  • User enters prompt in Unique AI ​

  • Prompt is within the ST on Azure

  • ​The prompt is sent to the LLM ​

  • If no web search is selected, only the LLM will respond based on the Data it was trained ​on

  • ​The system search queries derived from the prompt to the Google Search API

  • Crawler fetch the information via the links for websites that can be crawled​.

  • Crawlers only see the URL links. ​

  • Unique AI sends search results (snippets/URLs/notes) and context back to the LLM to generate a structured answer.​

  • The synthesized answer is returned to the user and rendered in the Unique AI UI. ​

Exposed Prompt elements

  • Internal: All prompt elements stay in the ST​

  • Internal: Prompt are personal to the user only admins with access can see. ​

  • Internal: All the prompt elements stay in the Single tenant. The LLM is within Azure. ​

  • External: Query terms sent if unsanitized: “Marie Dubois,” “Banque Genève”, “Audemars

  • Family”, “60 mio CHF”, “IT director responsibilities next 10 years”​

  • If sanitized: generic terms like “IT director asset management future responsibilities,” “case studies mid-size asset managers.”​

  • Internal: Any remaining sensitive context Unique AI includes (if original prompt or its identifiers are carried forward, they remain visible to the LLM)​

  • Internal: The original prompt and the model’s answer ​

  • In analytics/monitoring: partial prompt/response​

Data Retention and Location

  • Location: CH ​

  • Data Retention can be set manually

  • Azure backup defa​ult is 30 days

  • Location: CH

  • No data retention (Microsoft does not save any of the data) MS can't look at it and no training. ​

  • For Google, only data is sent to proximity-related search (e.g. priority given to CH and EU data processing) ​

  • ​Back to Switzerland​

  • Azure Tenant​

  • ​Back to Switzerland​

  • Azure Tenant​

Last updated