View Source Fnord

Tests | Dialyzer

Fnord is a command line tool the builds a searchable database of your files, using AI-generated embeddings to index and search your code base, notes, and other (non-binary) files. Notably, it provides a conversational interface to research within your project, answering complex questions about your code base.

Installation

  1. Install elixir if necessary:

    # MacOS
    brew install elixir
    
    # Debian-based
    sudo apt-get install elixir
    
  2. Add the mix escript path to your shell's PATH:

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

    mix escript.install github sysread/fnord
    

Use the same command to reinstall. It will offer to overwrite the existing installation.

  1. Set OPENAI_API_KEY

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.

Usage

Index

The first time you run this, especially on a large codebase, it will take a while to index everything. Subsequent runs will be faster, re-indexing only those files which have changed since they were last indexed.

fnord index --project foo --dir /path/to/foo

You can reindex the project, forcing it to reindex all files.

fnord index --project foo --dir /path/to/foo --reindex

You can also watch the project for changes and reindex them as they happen using watchman. Just be sure to use --quiet to suppress interactive output.

watchman-make -p '**/*' --settle 5 --run "fnord index --project $project --dir $project_root --quiet"

...or use the fnord-watch script in the tools directory on GitHub.

fnord-watch -p foo -d /path/to/foo

Search for files in the project that match a query.

fnord search --project foo --query "some search query"

If you want more detail about each file matched:

fnord search --project foo --query "some search query" --detail

Ask

Ask the AI assistant to answer questions about your project.

fnord ask foo "how do you run the tests for this project?"

# Pipe output to `glow` to render markdown
fnord ask foo "summarize the dependencies of this project" | glow

Miscellaneous

  • List projects: fnord projects
  • List files in a project: fnord files --project foo
  • Show the AI-generated summary of a file: fnord summary --project foo --file bar
  • Delete a project: fnord delete --project foo

Note that deleting a project only deletes from the index, not the actual files.

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)."
      },
      "concurrency": {
        "type": "integer",
        "description": "Number of concurrent threads to use for the search (default: 4)."
      }
    },
    "required": ["project", "query"]
  }
}

TODO

  • output formatted markdown results from ask
  • mitigate owl's bs; this thing is chincy af
  • we have GOT to come up with something bettern than Owl; it has SO many visual glitches, especially when there are many steps
  • restore progress display in indexer subcommand
  • nested tasks in tui
  • abstract agents out into a behaviour
    • support for tool calls
    • support for "accumulator" agents
  • add useful errors suggesting reindex when files are missing from the index