combfind 1.3.0

Queryable concept map of a codebase for LLM coding agents

combfind

Give an AI agent a codebase. combfind tells it where to look.

combfind builds a local index of a repository so an agent can find the right files and functions for a task with a plain-text query, without reading the entire codebase.

Install

For local LLM inference:

pip3 install "combfind[llm]" \
  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu

Download the default model (Qwen2.5-Coder-3B-Instruct Q6_K, ~2.5 GB):

combfind download-model

For a remote OpenAI-compatible API instead:

pip3 install "combfind[openai]"

For Apple Silicon (MLX):

pip3 install "combfind[mlx]"

Usage

# Index a repository (local LLM, auto-detected model)
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.*'

# Index using a remote 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

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

# Query it
combfind query "how does authentication work" --db repo.db
combfind query "where are database migrations" --db repo.db --format json

# Inspect a symbol returned by a query
combfind inspect auth.service.AuthService --db repo.db
combfind inspect auth.service.AuthService --db repo.db --format json

Query output (text)

[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

Query output (JSON)

[
  {
    "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": []
  }
]

Inspect output (text)

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

Init options

Flag	Default	Description
--db	<repo_path>/.combfind.db	Output path
--llm-model	auto-detected	Path to a GGUF model file (local mode only)
--llm-mode	local	LLM backend: local (llama.cpp), openai (OpenAI-compatible API), or mlx (Apple Silicon)
--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 LLM docstrings for undocumented symbols (slow)
--force	off	Re-run all stages, ignoring the cache

Query options

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

Inspect options

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

Environment variables

Variable	Default	Description
COMBFIND_LOG_LEVEL	info	Log verbosity: debug, info, warning, error
COMBFIND_MODEL	(auto-detected)	GGUF path for local mode / HF repo for mlx mode; equivalent to --llm-model
COMBFIND_LLM_BASE_URL	-	Base URL for OpenAI-compatible API (e.g. https://api.openai.com/v1)
COMBFIND_LLM_API_KEY	-	API key for the remote LLM
COMBFIND_LLM_MODEL	gpt-4o-mini	Model name to use with --llm-mode openai
HF_HUB_OFFLINE	-	Set to 1 to skip HuggingFace network checks and use cached embedding models only

Using a remote LLM API

Pass --llm-mode openai to use any OpenAI-compatible API instead of a local model. Configure it with environment variables:

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

Any API that speaks the OpenAI chat completions format works, including:

• OpenAI — set COMBFIND_LLM_BASE_URL=https://api.openai.com/v1
• Ollama — set COMBFIND_LLM_BASE_URL=http://localhost:11434/v1 and COMBFIND_LLM_API_KEY=ollama
• LM Studio — set COMBFIND_LLM_BASE_URL=http://localhost:1234/v1
• Any other OpenAI-compatible server — point COMBFIND_LLM_BASE_URL at its /v1 endpoint

--llm-model is ignored in openai mode; the model is selected via COMBFIND_LLM_MODEL.

Clustering

combfind groups symbols by their package/directory, then sub-clusters large packages using KMeans (targeting ~20 symbols per concept). This produces stable, interpretable concepts aligned with the codebase structure.

For best results, exclude test files at index time:

combfind init . --exclude-regex '.*test.*'

Supported languages

Python, Go, Java. More languages can be added via tree-sitter grammars.

Optional dependencies

These tools are not required but improve reference edge quality when installed:

Tool	Language	Install	Effect
scip-go	Go	go install github.com/scip-code/scip-go/cmd/scip-go@latest	Type-resolved call/import edges between Go symbols
scip-python	Python	npm install -g @sourcegraph/scip-python	Type-resolved call/import edges between Python symbols
scip-java	Java	See scip-java releases	Type-resolved call/import edges between Java symbols

Without them, combfind falls back to tree-sitter heuristics for import edges, which is less precise but works out of the box.

Contributing

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