View Source Fnord

• Description
• Features
• Installation
• Getting Started
• Tool usage
• Copyright and License

Description

fnord is a command line tool that uses multiple LLM-powered agents and tools to provide a conversational interface to your codebase, notes, and other (non-binary) files.

It can be used to generate on-demand tutorials, playbooks, and documentation for your project, as well as to search for examples, explanations, and solutions to problems in your codebase.

Why fnord?

AI-powered tools are limited by to the data built into their training data. RAG (Retrieval-Augmented Generation) using tool calls can supplement the training data with information, such as your code base, to provide more accurate and relevant answers to your questions.

But even with RAG, the AI still runs up against the context window. This is the conversational "memory" of the AI, often making use of an "attention mechanism" to keep it focused on the current instructions, but causing it to lose track of details earlier in the conversation.

If you've ever pasted multiple files into ChatGPT or worked with it iteratively on a piece of coce, you've probably seen this in action. It may forget constraints you defined earlier in the conversation or hallucinate entities and functions that don't exist in any of the files you've shown it.

fnord attempts to mitigate this will cleverly designed tool calls that allow the LLM to ask other agents to perform tasks on its behalf. For example, it can generate a prompt to ask another agent to read through a file and retrieve specific details it needs, like a single function definition, the declaration of an interface, or whether a specific functions behaves in a certain way. This keeps the entire file out of the "coordinating" agent's context window while still allowing it to use the information in the file to generate a response. This allows fnord to conduct more complex research across many files and directories without losing track of the details it needs to provide accurate answers.

Features

• Semantic search
• On-demand explanations, documentation, and tutorials
• Git archaeology
• Learns about your project(s) over time
• Improves its research capabilities with each interaction

Installation

Fnord is written in Elixir and is distributed as an escript.

• Install elixir

  # MacOS
  brew install elixir
  
  # Debian-based
  sudo apt-get install elixir
  

• Add the escript path to your shell's PATH

  echo 'export PATH="$HOME/.mix/escripts:$PATH"' >> ~/.bashrc
  source ~/.bashrc
  

• Install fnord

  mix escript.install github sysread/fnord
  

• Set OPENAI_API_KEY in your shell environment

Set this in your shell environment to the OpenAI project key you wish to use for this tool. You can create a new project and get a key here.

• Optional: Install a markdown viewer

Markdown seems to be the language of choice for LLMs, so installing something like glow to pipe output to will make the output more readable.

Getting Started

For the purposes of this guide, we will assume that fnord is installed and we are using it to interact with your project, blarg, which is located at $HOME/dev/blarg.

Index your project

The first step is to index your project to make it searchable with fnord's semantic search capabilities. The first time you index your project may take a while, especially if there are a lot of files. fnord respects .gitignore, but you may also wish to exclude certain paths or globs from indexing (e.g., -x 'node_modules').

fnord index --project blarg --dir $HOME/dev/blarg --exclude 'node_modules'

If you cancel partway-through (for example, with Ctrl-C), you can resume indexing by running the same command again.

You must re-index your project to reflect changes in the code base, new files, deleted files, or to change the directory or exclusions. This is done by running the same command again.

fnord stores its index in $HOME/.fnord/$project.

Search your code base

fnord's semantic search is powered by embeddings generated by OpenAI's text-embedding-3-large model. The indexing process does a lot more than simply generated embeddigs for the contents of your files. It also generates summary documentation for each file as well as an index of all functions, entities, and symbols in each file to enhance semantic matches for common questions and symbols.

fnord search --project blarg --query "callers of some_function"
fnord search --project blarg --query "some_function definition"
fnord search --project blarg --query "unit tests for some_function"

The summaries generated by the indexing process can also be included in the search results with the --detail flag.

fnord search --project blarg --query "some_function declaration" --detail | glow

If you would like to see more information about a single file, you can use the summary command.

fnord summary --project blarg --file "path/to/some_module.ext" | glow

Generate answers on-demand

fnord uses a combination of LLM-powered agents and tool calls to research your question within your project, including access to semantic search and git tools (read only!).

As it conducts its investigation, you will see some of the research steps reported back to you in real-time. These are printed to STDERR so they will not interfere with redirected output. If you wish to see more of this, set LOGGER_LEVEL=debug in your shell.

fnord ask --project blarg --question "Where is the unit test for some_function?"
fnord ask --project blarg --question "Find all callers of some_function"
fnord ask --project blarg --question "Is there a function or method that does X?"
fnord ask --project blarg --question "How do I add a new X implementation?"

Conversations (the transcript of messages between the LLM and the application) are saved for future reference and continuation. After each response, you will see a message like:

Conversation saved with ID: c81928aa-6ab2-4346-9b2a-0edce6a639f0

If desired, you can use that ID to continue the conversation with --follow.

fnord ask --project blarg --follow c81928aa-6ab2-4346-9b2a-0edce6a639f0 --question "Is some_function still used?"

List conversations with the conversations command. Adding --file will point to the file where the conversation is saved. --question will include the question that prompted the conversation.

fnord conversations --project blarg --question --file

As it learns more about your project, fnord will be able to answer more complex questions with less research.

Learning over time

fnord learns about your project over time by saving facts and inferences it makes about the project while researching your questions. These facts and inferences are saved and made searchable in the project's data in the $HOME/.fnord directory.

You can view the facts learned thus far, organized by topic, with the notes command.

fnord notes --project blarg | glow

Over time, these can become quite extensive, redundant, and stale over time as your code base evolves.

Consolidate and fact-check the notes for your project with the defrag command.

fnord defrag --project blarg

fnord knows how to "prime the pump" for its learning process with an initial set of learnings.

fnord ask --project blarg --question "Prime your research notes and report your findings"

Upgrades

fnord is under active development and new features are added regularly. To upgrade to the latest version, simply run:

fnord upgrade

Note that this is just a shortcut for:

mix escript.install github sysread/fnord

Other commands

• List projects: fnord projects
• List files in a project: fnord files --project foo
• Delete a project: fnord torch --project foo

Tool usage

Internally, the ask command uses the OpenAI chat completions API to generate a response, implementing a function tool to allow the assistant to query the database for information. fnord can be used to implement a similar tool for your own projects.

While the ask command severely limits the parameters that the assistant may utilize (query only, with project being provided by the user's invocation of the command), the following syntax includes the full set of parameters available for the search command.

{
  "name": "search_tool",
  "description": "Searches for matching files and their contents in a project.",
  "parameters": {
    "type": "object",
    "properties": {
      "project": {
        "type": "string",
        "description": "Project name for the search."
      },
      "query": {
        "type": "string",
        "description": "The search query string."
      },
      "detail": {
        "type": "boolean",
        "description": "Include AI-generated file summary if set to true."
      },
      "limit": {
        "type": "integer",
        "description": "Limit the number of results (default: 10)."
      },
      "workers": {
        "type": "integer",
        "description": "Number of concurrent workers to use for the search."
      }
    },
    "required": ["project", "query"]
  }
}

Copyright and License

This software is copyright (c) 2025 by Jeff Ober.

This is free software; you can redistribute it and/or modify it under the MIT License.