Backstory: We built Socratic after struggling to compile and maintain domain knowledge when building our own agents. At first, gathering all the relevant context from scattered docs and code to give the agent a coherent understanding was tedious. And once the domain evolved (e.g. changing specs and docs), the process had to be repeated. Socratic started as an experiment to see if this process can be automated.
The Problem: Building effective vertical agents requires high-quality, up-to-date, domain-specific knowledge. This is typically curated manually by domain experts, which is slow, expensive, and creates a bottleneck every time the domain knowledge changes.
The Goal: Socratic aims to automate this process. Given a set of unstructured source documents, Socratic identify key concepts, study them, and synthesize the findings into prompts that can be dropped directly into your LLM agent’s context. This keeps your agent's knowledge up-to-date with minimal overhead.
How it works: Given a set of unstructured domain documents, Socratic runs a lightweight multi-agent pipeline that: 1. Identifies key domain concepts to research. 2. Synthesizes structured knowledge units for each concept. 3. Composes them into prompts directly usable in your vertical agent’s context.
Socratic is open source and still early-stage. We would love your thoughts/feedbacks!