Langchain

Langgraph helped build Super by providing a standardized framework that made the complex LLM/source interactions more manageable through a declarative graph structure. It enabled better parallelization which significantly improved response speed, while also providing full graph state/sequence debugging capabilities that made troubleshooting easier. The migration to Langgraph unlocked new capabilities like human-in-the-loop pauses for clarification and better fallback mechanisms when models fail. Additionally, Langgraph's functional API allowed developers to write more natural, sequential code for complex subgraphs while still maintaining the benefits of the graph framework.

LangGraph's graph-based architecture with nodes and edges was the key on how we coordinate complex AI workflows. We're orchestrating multi-agent conversations where each node handles specific tasks (analysis, decision-making, action execution) and edges manage the flow between them. Memory lets us build operations that maintain context across conversations, while parallel execution nodes handle multiple AI operations simultaneously. The Pydantic formatted outputs ensure type-safe data flow between nodes, and the checkpoint system with PostgreSQL persistence lets us maintain state across these graph executions, so conversations can resume seamlessly even after server restarts.

We used langchain framework both for it's library agnostic methods for various AI platforms, as well as various helpers such as document parsers, chunking utilities, e.t.c. Moreover it has great agentic development support with langgraph. As an alternative we could have used OpenAI APIs directly however due to its fast changing nature, and also because we don't plan to depend on it long-term, it was better to go with langchain.