Convert notebooks to Markdown when specified in @file

[dlqqq]

Problem

Jupyter Notebooks (.ipynb) are serialized to disk as a JSON object. However, this object is not in a natural language format preferable to an LLM and includes various metadata that should not be sent to the LLM. When a notebook is passed to an LLM as a JSON object, this may result in lower-quality responses to questions about notebooks, as the contents of the cells are nested within the JSON object and are surrounded by irrelevant metadata.

Proposed Solution

When a notebook is specified in @file, the notebook should be converted to a Markdown string before being injected into the prompt sent to the LLM.

[JasonWeill]

Might be a good use for jupytext?

[michaelchia]

.ipynb files are not provided as raw json and are already handled as a special case as per

jupyter-ai/packages/jupyter-ai/jupyter_ai/context_providers/file.py

Lines 96 to 105 in 7c3d3a9

	return FILE_CONTEXT_TEMPLATE.format(
	filepath=filepath,
	content=self._process_file(content, filepath),
	)
	def _process_file(self, content: str, filepath: str):
	if filepath.endswith(".ipynb"):
	nb = nbformat.reads(content, as_version=4)
	return "\n\n".join([cell.source for cell in nb.cells])
	return content

.

However, is still a good idea to translate it to markdown, or at least handle/indicate different cell types correctly.