Skip to main content

Log - Sprint 7 🛫

· 3 min read

Work in Progress​

  • The team continue using Trello boards to track the cards in progress and using Slack to communicate blockers.
  • Today the team continued working and made progress on the following streams
    1. Multilingual capability:
      • The team observed that it was tricky to have a single collection in WD with two different languages, as WD can treat collection only in a single language. To solve this challenge, the team created two different collections, one with english and other with spanish both are the web crawl data.
      • Another challenge the team faced was dynamically figuring out which collection to be pointed to fetch results depending on question asked. Ideally if question is asked in spanish we need to fetch it from spanish collection in WD. To handle this challenge, the team developed the following prompt and gave instructions to the llama model to take care of mixed language and answer cleverly.
        • You need to deal with two languages that are English and Spanish. If the question is asked in English, then please answer in English. If the question is asked in Spanish, please answer back in spanish. Context for the question can be in Spanish or English or both the languages.
      • The team is also evaluating setting the first action to prompts the user to input their desired language
    2. FAQs to Actions: the team continue adding FAQs as actions. The next steps would be leveraging the watson assistant v2 api to create a python utility that has the following flow automated
      • set up the Assistant
      • transfer actions from one assitant to another assistant
      • WatsonX FAQs to actions
    3. High value journeys: the team finished developing the following actions to walk through users through a multi-step action
      • account opening
      • i lost a debit card
    4. Setup and integration of Watsonx Assistant and Genesys: The team continue the integration of Genesys to Watson Assistant, referring Integrating with Genesys. At around 12:30pm the team developed a static website hosted on COS that successfully integrate Watson Assistant with Genesys.
    5. Conversational Search Capability: For the questions that are not covered, GenAI conduct a search through the document collection. We tested token amounts to ensure comprehensiveness and conciseness of response
      • currently we are using the meta-llama/llama-2-70b-chat model, with max_new_token 300, repetition_penalty 1.1, and temperature 0.
    6. Website development using the Lendyr codebase: the team plan to integrate our Watson Assistant with the Lendyr codebase and customize it to match our client's information. This will provide the capabilities such as carousel, Map iframe, Journey, Email deep link, Mortgage Calculator, on top of the LLM Search.
      • at around 6pm, the team got the integration between the Lendyr codebase to work with Genesys integration

Currently Tracking​

  • The team is working on pushing solution documentation to GitHub.

Next Steps​

  • The team will continue to work through the use cases and demos.