PyCon Hong Kong 2025

So, you want to build an AI-powered application in Python. You will need a UI to accept user input, a database to store user data, and LLM deployments. You will also need to engineer prompts that guide the model to generate the desired output for your application. Yet, with all of these elements in place, there remains the immense challenge of coordinating their interaction. For a very simple task, filling a few user-inputted variables into a prompt for use with a single LLM deployment endpoint, the code for this might be only a few dozen lines in a single script. But what if you want to be able to swap between different models,. deployments, or APIs? And how will you handle problems such as hitting quota limits or encountering irregularities in LLM output, or test how you process inputs and outpus? Or what if you want to build a more complex AI system, such as one that draws from internal or external data sources to enrich the model's context (RAG) or that is able to dynamically generate additional prompts in response to user requests (agents)? These are all questions that concern how to build an AI pipeline. Several Python Python packages -- including LangChain, LlamaIndex, and Haystack -- claim to offer end-to-end frameworks for building sophisticated AI pipelines. The purpose of this talk is to present a systematic comparison of the alternatives available to developers working in Python for AI pipeline construction, including the option of just writing a framework for oneself.