With AI increasingly automating away code generation, software engineers will spend more time reading, judging, and architecting code rather than writing it. Storia is building an open-source copilot that knows a company’s codebase and its context.
We are starting with Sage, a Perplexity-like agent for helping developers understand, judge, and generate software. Given an existing codebase, developers can ask Sage questions such as:
1) Given my project’s SLA and latency constraints, what is the appropriate underlying vector database to use? How would I incorporate it into my existing codebase?
2) Why should I pick Redis over Milvus as my underlying vector store?
3) Does this codebase in our organization still work and what steps are required for a complex integration with another library?
Sage’s answers are directly supported by documentation and external references like GitHub, Stack Overflow, technical design documents, and project management software, preventing hallucinations. Today, Sage has up-to-date knowledge about open-source repositories (indexed daily). Tomorrow it will have a deep understanding of every line of code on the Internet. For teams, Sage will know about your private codebase too.
No group has yet solved how to build an AI system that comprehends a codebase and its context and can empower every developer to architect better code, faster. This requires new research advances because vanilla RAG and out-of-the-box LLMs aren’t going to cut it.
We have 20+ years of software engineering and AI research experience. Julia worked on precursors of Gemini using contextual neural techniques before they were called “RAG” (and applied it to products like Google Keyboard and Pixel phones). Mihail built the earliest LLMs at Amazon Alexa and launched the first contextual deep learning conversational AI system in production at Alexa.