Understood: AI Assistant MVP

The Challenge

Understood.org reaches millions of parents seeking support for children with learning and thinking differences. While they have an extensive library of peer-reviewed articles and clinical advice, parents found it difficult to locate highly specific answers to their daily struggles. Understood needed a smart, conversational assistant that could answer questions using only their verified medical corpus, preventing OpenAI models from hallucinating false details.

Our Solution

vector software inc deployed a product squad to engineer a Retrieval-Augmented Generation (RAG) system. We first created data ingestion jobs to chunk and store Understood's article assets into vector embeddings in Pinecone.

Key Technical Decisions

To optimize response accuracy and manage running costs, our engineering pod made several critical architectural choices:

• Hybrid Vector-Keyword Indexing: We combined Pinecone vector semantic search with a traditional Elasticsearch keyword filter, ensuring highly specific medical terms (like clinical diagnoses) resolve perfectly.
• Prompt Evaluation Testing: We built an offline test runner that tests prompt updates against 200 standard user queries, automatically rating performance to avoid regression.
• Edge-Caching Layer: Repeated queries are cached via a Redis database cluster, reducing LLM runtime API costs by over 45% during peak hours.

The Impact

Within 12 weeks, we transitioned the solution from a whiteboard blueprint to a production-ready MVP. In early beta tests with thousands of active parents, the AI assistant achieved a 94% relevance rating, with search times dropping from minutes to seconds.