Skip to content

Guide to examples in langroid-examples repo

The langroid-examples repo contains several examples of using the Langroid agent-oriented programming framework for LLM applications. Below is a guide to the examples. First please ensure you follow the installation instructions in the langroid-examples repo README.

At minimum a GPT4-compatible OpenAI API key is required. As currently set up, many of the examples will not work with a weaker model. Weaker models may require more detailed or different prompting, and possibly a more iterative approach with multiple agents to verify and retry, etc — this is on our roadmap.

All the example scripts are meant to be run on the command line. In each script there is a description and sometimes instructions on how to run the script.

NOTE: When you run any script, it pauses for “human” input at every step, and depending on the context, you can either hit enter to continue, or in case there is a question/response expected from the human, you can enter your question or response and then hit enter.

Basic Examples

  • /examples/basic/chat.py This is a basic chat application.

    • Illustrates Agent task loop.

  • /examples/basic/autocorrect.py Chat with autocorrect: type fast and carelessly/lazily and the LLM will try its best to interpret what you want, and offer choices when confused.

    • Illustrates Agent task loop.

  • /examples/basic/chat-search.py This uses a GoogleSearchTool function-call/tool to answer questions using a google web search if needed. Try asking questions about facts known after Sep 2021 (GPT4 training cutoff), like when was llama2 released

    • Illustrates Agent + Tools/function-calling + web-search

  • /examples/basic/chat-tree.py is a toy example of tree-structured multi-agent computation, see a detailed writeup here.

    • Illustrates multi-agent task collaboration, task delegation.

Document-chat examples, or RAG (Retrieval Augmented Generation)

  • /examples/docqa/chat.py is a document-chat application. Point it to local file, directory or web url, and ask questions
    • Illustrates basic RAG
  • /examples/docqa/chat-search.py: ask about anything and it will try to answer based on docs indexed in vector-db, otherwise it will do a Google search, and index the results in the vec-db for this and later answers.
    • Illustrates RAG + Function-calling/tools
  • /examples/docqa/chat_multi.py: — this is a 2-agent system that will summarize a large document with 5 bullet points: the first agent generates questions for the retrieval agent, and is done when it gathers 5 key points.
    • Illustrates 2-agent collaboration + RAG to summarize a document
  • /examples/docqa/chat_multi_extract.py: — extracts structured info from a lease document: Main agent asks questions to a retrieval agent.
    • Illustrates 2-agent collaboration, RAG, Function-calling/tools, Structured Information Extraction.

Data-chat examples (tabular, SQL)

  • /examples/data-qa/table_chat.py: - point to a URL or local csv file and ask questions. The agent generates pandas code that is run within langroid.
    • Illustrates function-calling/tools and code-generation
  • /examples/data-qa/sql-chat/sql_chat.py: — chat with a sql db — ask questions in English, it will generate sql code to answer them. See tutorial here
    • Illustrates function-calling/tools and code-generation