combfind 1.5.0

Queryable concept map of a codebase for LLM coding agents

combfind

When an AI coding agent gets a ticket like "users get logged out randomly on mobile," it has two failure modes: it reads too many files burning tokens and time, or it finds a relevant file and patches it locally, missing that the bug lives in shared code, an interface, or a sibling implementation.

combfind fixes this. It builds a concept map of a codebase so an agent can query "session token refresh" and get back ranked symbols with files and line ranges. The key is what it tells you about structure: is this an interface, an implementation, or one of several siblings that all need to change together? That context is what prevents a local patch to the wrong layer. In practice it cuts orientation-phase token cost by 50-66% (measured on one dev loop; your mileage will vary): the agent reads 3-5 targeted files instead of scanning dozens.

Runs entirely locally. Doesn't require paid APIs.

Install

# Local LLM (llama.cpp)
pip install "combfind[llm]" \
  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu

# Apple Silicon (MLX)
pip install "combfind[mlx]"

# Remote OpenAI-compatible API
pip install "combfind[openai]"

# Gleam support
pip install "combfind[gleam]"

Download the default local model (~2.5 GB):

combfind download-model

Quick start

# Build the index
combfind init /path/to/repo --db repo.db

# Query it
combfind query "how does authentication work" --db repo.db

# Inspect a symbol from the results
combfind inspect auth.service.AuthService --db repo.db

Usage

init: build the index

# Basic
combfind init /path/to/repo --db repo.db

# Exclude test files (recommended for cleaner concepts)
combfind init /path/to/repo --db repo.db --exclude-regex '.*test.*'

# OpenAI-compatible API
COMBFIND_LLM_API_KEY=sk-... COMBFIND_LLM_MODEL=gpt-4o-mini \
  combfind init /path/to/repo --db repo.db --llm-mode openai

# Apple Silicon MLX
combfind init /path/to/repo --db repo.db --llm-mode mlx \
  --llm-model mlx-community/Qwen2.5-7B-Instruct-4bit

Flag	Default	Description
--db	<repo_path>/.combfind.db	Output database path
--llm-mode	local	LLM backend: local, openai, or mlx
--llm-model	auto-detected	GGUF path (local) or HF repo ID (mlx)
--exclude-paths		Paths to skip, relative to repo root (repeatable)
--exclude-regex		Regex matched against file paths to skip
--llm-workers	1	Parallel LLM calls (useful with --llm-mode openai)
--docgen	off	Generate docstrings for undocumented symbols (slow)
--force	off	Re-run all stages, ignoring the cache

query: search the index

combfind query "users get logged out randomly" --db repo.db
combfind query "where are database migrations" --db repo.db --format json

Text output:

[1] Token Refresh (implementation) - 0.87
    why: Handles session token validation and refresh logic.
    auth/service.py
      auth.service.AuthService.refresh  :42-67
      auth.service.AuthService.validate  :70-91

JSON output:

[
  {
    "rank": 1,
    "concept": "Token Refresh",
    "role": "implementation",
    "score": 0.87,
    "files": [
      {
        "path": "auth/service.py",
        "symbols": [
          {"name": "refresh", "qualified_name": "auth.service.AuthService.refresh", "start_line": 42, "end_line": 67},
          {"name": "validate", "qualified_name": "auth.service.AuthService.validate", "start_line": 70, "end_line": 91}
        ]
      }
    ],
    "why_relevant": "Handles session token validation and refresh logic.",
    "sibling_implementations": []
  }
]

Flag	Default	Description
--db	.combfind.db	Database to query
--top-k	5	Number of results
--format	text	text or json
--rerank	off	Re-score results with LLM (requires --llm-mode)
--agentic	off	Iterative query loop: LLM steers follow-up searches until satisfied (requires --llm-mode)
--agentic-limit	3	Max iterations for --agentic
--llm-mode		LLM backend for --rerank / --agentic: local, openai, or mlx

inspect: look up a symbol

combfind inspect auth.service.AuthService --db repo.db
combfind inspect auth.service.AuthService auth.service.TokenService --db repo.db --format json

Output:

auth.service.AuthService  (class, auth/service.py:10-80)
concept:  Token Refresh  [implementation]
sig:      class AuthService

callers (1):
  auth.mock.MockAuthService  auth/mock.py:5

callees (1):
  auth.service.AuthService.validate  auth/service.py:20

concept siblings (1):
  auth.service.AuthService.validate  [method]  auth/service.py

Flag	Default	Description
--db	.combfind.db	Database to query
--format	text	text or json

How it works

The init pipeline runs six stages, each reading and writing to a SQLite file:

1. parse: tree-sitter extracts files, symbols (signatures, line ranges, docstrings, imports)
2. index: SCIP or tree-sitter heuristics populate a references table of calls, imports, and inheritance edges
3. embed: sentence-transformers produces a vector per symbol
4. cluster: symbols are grouped by package/directory, then sub-clustered with KMeans (~20 symbols per concept)
5. label: a local LLM names and describes each cluster and assigns a structural role (interface | implementation | orchestrator | entry_point | domain_model | infrastructure | cross_cutting)
6. embed concepts: sentence-transformers produces a vector per concept description

At query time: embed the query, cosine search over concept embeddings, optionally rerank with LLM, expand top concepts to member symbols and 1-hop callers/callees, return ranked symbols and code regions.

Stages are cached by a content hash of their inputs. When you re-run init, only stages affected by changed files are re-executed; the rest are skipped. Pass --force to rebuild from scratch.

Performance

On a 50k LOC Go codebase using Qwen2.5:7b via Ollama, the initial index builds in ~5 minutes. Query time is around 7 seconds, most of which is loading the local model.

Incremental reindexing is fast. When a handful of files change, re-running init takes around 30 seconds; only the stages affected by changed files are re-executed. The index is also crash-safe: progress is committed to SQLite in batches within each stage, so if a run is interrupted it picks up close to where it left off rather than starting over.

The goal is not to replace careful code reading. It is to give an agent a cheap orientation pass so it knows which 3-5 files to read rather than all 500. On that goal, combfind achieves file_recall@3 of 0.75 on structural queries with --rerank, evaluated against 10 real bug fixes from a production Go codebase. That puts it above dense retrieval baselines like BM25 and E5-large (NDCG ~0.57-0.59 per Practical Code RAG at Scale, 2025), with no API costs. The state of the art (Agentless with frontier models) reaches ~90% recall@5, but requires expensive multi-step LLM pipelines per query. combfind trades some accuracy for being fast, cheap, and fully local.

How to query well

combfind matches against concept descriptions, so structural queries outperform symptom descriptions.

"Where are user creation request DTOs and their field definitions?" finds the right code immediately. "EmailVerified boolean gets rejected by the validator" does not, because the symptom vocabulary has no overlap with the code structure.

When an agent receives a bug ticket, the right move is to translate the symptom into a structural question before querying: not what went wrong, but where does this kind of code live.

Supported languages

Python, Go, Java, Gleam, Erlang.

Optional SCIP tools

These are not required but produce more accurate call and import edges than the tree-sitter fallback:

Tool	Language	Install
scip-go	Go	go install github.com/scip-code/scip-go/cmd/scip-go@latest
scip-python	Python	npm install -g @sourcegraph/scip-python
scip-java	Java	scip-java releases

Using a remote LLM

Pass --llm-mode openai to use any OpenAI-compatible API:

export COMBFIND_LLM_BASE_URL=https://api.openai.com/v1
export COMBFIND_LLM_API_KEY=sk-...
export COMBFIND_LLM_MODEL=gpt-4o-mini

combfind init /path/to/repo --db repo.db --llm-mode openai

Works with OpenAI, Ollama (http://localhost:11434/v1), LM Studio (http://localhost:1234/v1), and any other OpenAI-compatible server.

Environment variables

Variable	Default	Description
COMBFIND_LOG_LEVEL	info	Log verbosity: debug, info, warning, error
COMBFIND_MODEL	auto-detected	GGUF path (local) or HF repo ID (mlx); equivalent to --llm-model
COMBFIND_LLM_BASE_URL		Base URL for OpenAI-compatible API
COMBFIND_LLM_API_KEY		API key for remote LLM
COMBFIND_LLM_MODEL	gpt-4o-mini	Model name for --llm-mode openai
HF_HUB_OFFLINE		Set to 1 to use cached embedding models without network access

Contributing

See CONTRIBUTING.md for dev setup, commit conventions, and the release pipeline.