Series · 5 parts · December 2024 – January 2025

Building a Retrieval Agent with LangGraph

I wanted to know what building an AI solution with LangGraph actually feels like, so I took the retrieval-agent template and pushed it the whole way: jump start from a template, extend it, debug it, deploy it — and then keep fixing what real use exposed.

The arc, in one breath

Part 1 takes LangGraph's RAG template and makes it mine — crawling from a list of starter URLs within a set number of hops, indexed into Milvus. Part 2 opens the graph up to a debugger, so the thing can be inspected while it runs rather than guessed at. Part 3 puts it on LangGraph Cloud, and Part 4 gives it a face by pointing an assistant-ui chat client at the deployment. Then Part 5 is the part the template never warns you about: crawls that time out, retrieval that is too coarse, embedding rate limits, and demos too slow to give.

Part 1 · The Template · December 31, 2024

Expanding the LangGraph Retrieval Agent

Start from LangGraph's RAG template and extend it: crawl from a list of starter_urls within a given number of hops to build the vector index, and swap in the open-source Milvus as the retriever.

Read Part 1 →

Part 2 · Debugging · January 1, 2025

Step-thru Debugging your LangGraphs

Run the API server with --debug-port, attach debugpy from VS Code, and put a breakpoint on any graph node to step through execution instead of inferring it from the outside.

Read Part 2 →

Part 3 · Deployment · January 6, 2025

Deploying to LangGraph Cloud

Import the repo straight from GitHub, add the environment configuration, submit — then wait out the 15-minute deploy, which you pay again in full for every environment-variable change.

Read Part 3 →

Part 4 · The Chat Client · January 8, 2025

Adding a Chat Client to a LangGraph Assistant

Point an assistant-ui chat client at the deployed graph: scaffold from the LangGraph template, create an assistant in LangGraph Studio, and wire the two together with three environment variables.

Read Part 4 →

Part 5 · What Real Use Exposed · January 16, 2025

Improvements to the Retrieval Agent

Four fixes: loading pre-crawled Apify datasets when crawling would time out, splitting documents for finer-grained retrieval, batching around embedding rate limits, and pre-created indexes so demonstrations stay fast and cheap.

Read Part 5 →