Stealth financial start-up has an existing rudimentary MVP for a long-form AI chatbot with RAG for internal knowledge base. The system required huge improvements for the next phase of development. Specifically, we were tasked to improve the system in these forms: response relevance, information accuracy, better long-term and accurate conversation memory.  

We initiated with a Discovery and Design phase: 

1) Audited their business to better understand the goals of the project, value propositions and technical requirements, KPIs, and financial or time constraints. 

2) Provided technical design, milestones, deliverables, existing data analysis, etc… 

3) Designed key success metrics and automated response evaluation AI models for systematic quantification of response performance

The solution we provided had improved the MVP in these areas:

Response Relevance: Improved prompt engineering using few-shot prompting and chain of thoughts.

Retrieval accuracy: RAG model was improved with Multivector Retrieval (improve semantic matching of relevant documents) and Re-ranking of retrieved chunks (improves ordering of most relevant contents)

Long-term complex memory: Multi-dimensional chat memory map-reduce summarization to summarize extended conversations into various salient information dimensions. This is then stored and restored using RAG for long-term use.