Skip to content

langchain-ai/kork

Repository files navigation

Unit Tests

NOT MAINTAINED

Kork was a prototype for agents that can take multiple (dependent) actions at once; i.e., when the input of action 2 depends on the output from action 1.

This is distinct from parallel tool use where the inputs into the tools are independent of the tools outputs.

We're currently not maintaining Kork, but you're welcome to use it for your own research.

Kork alt The Parrot

Kork is an experimental Langchain chain that helps build natural language APIs powered by LLMs.

Features

  1. Assemble a natural language API from a set of python functions.
  2. Generate a prompt to help the LLM write a correct program.
  3. Execute the program generated by the LLM safely.
  4. Program generation takes a single LLM call, as opposed to agents that use a call per action taken.

Show me! 😹

Image manipulation API using PIL

Cats!!!

How

The chain takes a user query, translates it into a program, and executes it using the Kork interpreter.

The interpreter is used to run code written in the Kork programming language.

The language is limited by design to variable declarations, function invocations and arithmetic operations.

This limitation allows controlling the kind of programs the LLMs can generate, making the output of the LLM more predictable and safer to execute in a production setting. (An agent with access to "constrained" tools has similar benefits, but requires a call per action taken.)

Kork adds a few additional tricks (e.g., retriever interfaces for examples and foreign functions) to help guide the LLM to generate a correct program.

What?

No loops, no conditionals, no file access, no network access, no arbitrary code execution. WHAT?!

The ability to invoke custom functions goes a long way in terms of the kinds of programs that can be written! (You can always add a write_to_file function!)

Limitations

  • Kork cannot write arbitrary code. If that's what you need, save yourself some time and use docker and a real programming language.
  • The Kork chain is not an Agent, so it cannot inspect intermediate steps, instead it generates a pre-defined set of function calls that need to be executed. (But remember that function calls can involve calls to an agent!)
  • The Kork language and interpreter are limited to function invocation, variable declaration and basic arithmetic (no function declaration, loops etc.)
  • Only supporting int, float, str, type(None), bool types. No support for lists or object types.
  • Very limited type annotations.

Quality

The short version: No benchmarks yet!

The long version: Quality of generated programs depends on many factors.

You can experiment with tweaking the prompt (e.g., explain the syntax of the language), changing the foreign function retriever (e.g., retrieve the most relevant foreign functions based on the user query), or providing examples in the form of (query, expected program).

The prompt, examples and syntax of the language can trip up the LLM to assume it's programming in a specific language (e.g., typescript or python) and assume that it can use language features that are not supported by Kork or to import libraries that are not available.

Safety

Nothing is bulletproof when humans are involved.

  • Don't do silly things like exposing eval as a foreign function.
  • If a foreign function allocates memory, the LLM could ask for more memory than is available crashing the process.
  • Think carefully before using the output of an LLM. If the LLM behaves maliciously, is it safe to use an LLM in the given context?

Future Work

Let us know if you're interested in contributing or have ideas for improvements!

  • Allow enabling/disabling language features
  • Allow changing underlying language syntax
  • Add support for objects
  • Add other language features (e.g., loops)
  • Provide foreign function retriever implementations based on similarity to user query

Compatibility

Kork is tested against python 3.8, 3.9, 3.10, 3.11.

Installation

pip install kork

🙏 Appreciation

© Contributing

If you have any ideas or feature requests, please open an issue and share!

See CONTRIBUTING.md for more information.

🎶 Why the name?

Fast to type and maybe sufficiently unique.