zed-zeta-bindings

Python bindings for Zed's Zeta prompt formatting and model output parsing code.

Installation

pip install zed-zeta-bindings

Local builds

Local builds need a Zed checkout because this package depends on Zed's crates/zeta_prompt and crates/edit_prediction_metrics crates. The build helpers generate a gitignored .zed-path/Cargo.toml manifest that points Cargo at the discovered checkout.

Set ZED_PATH to the checkout root when you want to be explicit:

ZED_PATH=/path/to/zed uv run zeta bindings build --native

When ZED_PATH is unset, the helpers try common local layouts including ../zed, ~/zed/zed, ~/zed, and ~/develop/zed.

Usage

import zeta_bindings as zeta

prompt_format = "V0318SeedMultiRegions" # Zeta2.1 prompt format
prompt_input = zeta.ZetaPromptInput.from_dict(request_body)
prompt = zeta.format_prompt(prompt_input, prompt_format)
stop_tokens = zeta.stop_tokens(prompt_format) or None

raw_output = await engine.predict({"prompt": prompt, "stop": stop_tokens})
parsed_output = zeta.parse_output(raw_output, prompt_format, prompt_input)

response = {
    "output": parsed_output.new_editable_region,
    "editable_range": {
        "start": parsed_output.range_in_excerpt[0],
        "end": parsed_output.range_in_excerpt[1],
    },
}

For more detailed type information, see crates/zeta_prompt in the Zed repository: https://github.com/zed-industries/zed/tree/main/crates/zeta_prompt.

License

zed-zeta-bindings is licensed under GPL-3.0-or-later.

API

Functions that accept a format_str parameter are parameterized on the prompt format. Refer to the following table for the prompt format to use:

ZetaPromptInput.from_dict(data)

Builds the typed prompt input from a request dictionary. The required fields are:

• cursor_path
• cursor_excerpt
• cursor_offset_in_excerpt
• excerpt_ranges

Optional fields include excerpt_start_row, events, related_files, active_buffer_diagnostics, and syntax_ranges.

format_prompt(prompt_input, format_str)

Formats a ZetaPromptInput into the prompt string sent to the model.

Returns None if the input does not fit within the selected format's token budget.

stop_tokens(format_str)

Returns the stop tokens for the selected prompt format.

parse_output(model_output, format_str, prompt_input)

Parses a raw model completion into a ParsedOutput.

ParsedOutput contains:

• new_editable_region: the replacement text for the editable region.
• range_in_excerpt: the [start, end] byte range of the editable region in cursor_excerpt.
• cursor_offset_in_new_editable_region: the cursor offset in the replacement text, when present.