Skip to content

Latest commit

 

History

History
56 lines (50 loc) · 3.16 KB

File metadata and controls

56 lines (50 loc) · 3.16 KB

Future Extensibility 

The current solution focuses mainly on post-call audio and text conversation knowledge mining. There are a number of additional capabilities and considerations that would complement this accelerator and are areas for extending: 

  1. Topic modeling

    Key phrases and complaints are generated by LLM, however, since the LLM processes only one conversation at a time, this often leads to unintended duplicates (e.g., flight delay = delayed flight = flight delayed). Topic modeling could be implemented to handle this more elegantly.

  2. Improved system of aggregate insight prompts 

    The accelerator utilizes system prompts that are used to derive insights from a conversation. These are customizable within the accelerator notebooks, but additional insights may be required based on new use cases for this accelerator, which will require changes to the existing prompts or additional prompts to be created. In addition, new techniques or approaches to the prompt structure may be made to enhance the accuracy and relevancy, especially as new versions of GPT are released.

  3. Q&A style interface for users to ask questions across all conversations

    In the current accelerator, insights are surfaced by using filters in a Power BI dashboard. Using a Q&A approach would unlock a conversational approach and follow up questions. This could also be extended to allow users to derive insights on their own, without pre-processing across the full dataset. 

  4. Historical reprocessing workflow for all conversations to aggregate new insights as they're authored 

    It's common for additional insights or enhancements to prompts to be made over time, but as these changes are made the existing conversations will need to be reprocessed. Currently, the solution does not have a "reprocess" capability and would require the conversations to be removed and then readded. For ease of use, a reprocessing capability could be added to simply rerun the skills on the index either on a regular cadence, or manually running across the full index. This ensures most up-to-date insights and capabilities within the mined knowledgebase. 

  5. Video and other conversation formats 

    The solution only accepts text and audio based conversations, but this could be extended to support additional formats as well -- including video and images sets. This extends the ability to support additional channels customer service and users interact through, still mining the same knowledge insights into a single system. 

  6. Opt-out flag on conversations to skip knowledge mining processing

    Privacy and legal requirement around data processing is a consideration that all data processing platforms should consider. This solution can be extended to support an explicit parameter or flag on conversations that users have opted out or in to their conversation being processed by AI. The solution can then decide to skip over specific conversations where users have opted out.