Build a Docstring Search Engine

In this tutorial you’ll build a semantic search tool on top of kit using docstring-based indexing.

Why docstrings? Summaries distill intent rather than syntax. Embedding these short natural-language strings lets the vector DB focus on meaning, giving you relevant hits even when the literal code differs (e.g., retry() vs attempt_again()). It also keeps the index small (one embedding per file or symbol instead of dozens of raw-code chunks).

---

1. Install dependencies

uv pip install kit[openai] sentence-transformers chromadb

2. Initialise a repo and summarizer

import kit
from kit import Repository, DocstringIndexer, Summarizer, SummarySearcher
from sentence_transformers import SentenceTransformer

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

summarizer = repo.get_summarizer()  # defaults to OpenAIConfig

3. Build the docstring index

embed_model = SentenceTransformer("all-MiniLM-L6-v2")
embed_fn = lambda txt: embed_model.encode(txt).tolist()

indexer = DocstringIndexer(repo, summarizer, embed_fn)
indexer.build()          # writes REPO_PATH/.kit_cache/docstring_db

The first run will take time depending on repo size and LLM latency. Summaries are cached inside the vector DB (and in a meta.json within the persist_dir), so subsequent runs are cheap if code hasn’t changed.

4. Query the index

searcher = indexer.get_searcher()

results = searcher.search("How is the retry back-off implemented?", top_k=3)
for hit in results:
    print(f"→ File: {hit.get('file_path', 'N/A')}\n  Summary: {hit.get('summary', 'N/A')}")

You now have a semantic code searcher, using powerful docstring summaries, as easy as that.