Assembling Context

When you send code to an LLM you usually don’t want the entire repository – just the most relevant bits. ContextAssembler helps you stitch those bits together into a single prompt-sized string.

Why you need it

• Token limits – GPT-4o tops out at ~128k tokens; some models less.
• Signal-to-noise – Cut boilerplate, focus the model on what matters.
• Automatic truncation – Keeps prompts within your chosen character budget.

Quick start

from kit import Repository, ContextAssembler

repo = Repository("/path/to/project")

# Assume you already have chunks, e.g. from repo.search_semantic()
chunks = repo.search_text("jwt decode")

assembler = ContextAssembler(max_chars=12_000)
context = assembler.from_chunks(chunks)

print(context)  # → Ready to drop into your chat prompt

chunks can be any list of dicts that include a code key – the helper trims and orders them by length until the budget is filled.

Fine-tuning

Parameter	Default	Description
max_chars	12000	Rough character cap for the final string.
separator	"\n\n---\n\n"	Separator inserted between chunks.
header / footer	""	Optional strings prepended/appended.

assembler = ContextAssembler(
    max_chars=8000,
    header="### Code context\n",
    footer="\n### End context",
)

Combining with other tools

1. Vector search → assemble → chat

   chunks = repo.search_semantic("retry backoff", embed_fn, top_k=10)
   prompt = assembler.from_chunks(chunks)
   response = my_llm.chat(prompt + "\n\nQ: …")

2. Docstring search first – Use SummarySearcher for high-level matches, then pull full code for those files via repo.context.
3. Diff review bots – Feed only the changed lines + surrounding context.

API reference

from kit.llm_context import ContextAssembler

__init__(repo, *, title=None)

Constructs a new ContextAssembler.

• repo: A kit.repository.Repository instance.
• title (optional): A string to prepend to the assembled context.

from_chunks(chunks, max_chars=12000, separator="...", header="", footer="")

This is the primary method for assembling context from a list of code chunks.

• chunks: A list of dictionaries, each with a "code" key.
• max_chars: Target maximum character length for the output string.
• separator: String to insert between chunks.
• header / footer: Optional strings to wrap the entire context.

Returns a single string with concatenated, truncated chunks.

Other methods

While from_chunks is the most common entry point, ContextAssembler also offers methods to add specific types of context if you’re building a prompt manually:

• add_diff(diff_text): Adds a Git diff.
• add_file(file_path, highlight_changes=False): Adds the full content of a file.
• add_symbol_dependencies(file_path, max_depth=1): Adds content of files that file_path depends on.
• add_search_results(results, query): Formats and adds semantic search results.
• format_context(): Returns the accumulated context as a string.