Semantic Analysis

Semantic Analysis

The semantic pass enriches the graph with LLM-derived concepts and resolves ambiguous duplicates that the deterministic pipeline cannot. It is opt-in via codegraph extract --semantic, needs an LLM backend API key, and makes paid API calls (except for local/subscription backends).

codegraph extract . --semantic

When --semantic is set but no backend key is detected, or backend init fails, the pass is skipped with a note and extraction continues normally. The pass is never run by update, watch, or the git hooks; those preserve existing semantic nodes but do not generate new ones (see [Incremental-Updates]).

See also: [Extraction], [Configuration], [MCP-Server].

What the semantic pass does

With --semantic and a working backend, the extract run adds three LLM-driven steps on top of the deterministic build:

1. Concept extraction. Documents and papers (not code) are sent to the model, which returns concept nodes and edges. These are merged into the graph before clustering, so concepts can collapse onto the AST symbols they describe. Documents are chunked to a token budget (60,000 tokens per chunk) and, on context overflow or truncated output, the chunk is recursively bisected (up to depth 3) and the partial results merged.
2. Dedup tiebreaker. The deterministic dedup leaves "ambiguous" concept pairs (fuzzy label similarity in the 75-92 band) unmerged. With an LLM, the model is asked, in batches, whether each pair names the same real-world concept; confirmed pairs are merged. Offline (no LLM), a conservative deterministic rule merges only word-reorderings/duplications and flags the rest for review.
3. Community labeling. Communities are named by the model from their representative member labels. Without a backend, labels stay empty and views fall back to Community N.

The output reports what the LLM contributed, for example:

Semantic pass: using the claude backend
Semantic pass: +18 concept node(s), +24 edge(s)
Dedup tiebreaker (LLM): merged 3 of 7 ambiguous concept pair(s), 4 left for review
LLM usage (extraction pass): 41200 input + 5300 output tokens (~$0.2031 estimated on claude)

Supported LLM backends

CodeGraph selects one backend from the environment. Backends fall into three groups:

• OpenAI-compatible Chat Completions: OpenAI, Gemini (its OpenAI-compat layer), Kimi (Moonshot), DeepSeek, Azure OpenAI, and Ollama (local).
• Native Anthropic Messages API (POST /v1/messages).
• AWS Bedrock (Converse API, SigV4-signed; AWS env credentials only).
• claude CLI (the locally-installed Claude Code CLI; opt-in only).

Selection and priority

The backend is resolved from environment variables in this priority order: gemini, kimi, claude, openai, deepseek, azure, bedrock, ollama. The first one whose credentials are present wins. Set CODEGRAPH_BACKEND to force a specific backend by name (gemini, kimi, claude, openai, deepseek, azure, bedrock, ollama, claude-cli). claude-cli is never auto-detected; CODEGRAPH_BACKEND=claude-cli is the only way to select it.

Backends and their environment variables

Backend	API key / credential	Model override (env)	Base URL / endpoint override (env)	Default model
gemini	GEMINI_API_KEY or GOOGLE_API_KEY	GEMINI_MODEL	(built in)	gemini-2.5-flash
kimi	MOONSHOT_API_KEY	MOONSHOT_MODEL	(built in)	kimi-k2
claude	ANTHROPIC_API_KEY	ANTHROPIC_MODEL	ANTHROPIC_BASE_URL	claude-sonnet-4-6
openai	OPENAI_API_KEY	OPENAI_MODEL	OPENAI_BASE_URL	gpt-4.1-mini
deepseek	DEEPSEEK_API_KEY	DEEPSEEK_MODEL	(built in)	deepseek-chat
azure	AZURE_OPENAI_API_KEY + AZURE_OPENAI_ENDPOINT	AZURE_OPENAI_DEPLOYMENT	AZURE_OPENAI_ENDPOINT	gpt-4o
bedrock	AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY	BEDROCK_MODEL	(region endpoint)	anthropic.claude-3-5-sonnet-20241022-v2:0
ollama	(keyless; OLLAMA_API_KEY optional)	OLLAMA_MODEL	OLLAMA_BASE_URL	qwen2.5-coder:7b
claude-cli	(uses your claude CLI login)	CLAUDE_CLI_MODEL	n/a	(CLI default)

Notes:

• Azure requires both a key and an endpoint. It addresses the model as a deployment path with an api-key header; AZURE_OPENAI_DEPLOYMENT is the deployment name, and AZURE_OPENAI_API_VERSION overrides the REST API version (default 2024-12-01-preview).
• Ollama is local and keyless; setting OLLAMA_BASE_URL opts it in. It runs serially.
• Bedrock uses AWS environment credentials only (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, optional AWS_SESSION_TOKEN, and AWS_REGION / AWS_DEFAULT_REGION, default region us-east-1); AWS profile and instance-role resolution are not supported.
• CODEGRAPH_LLM_TEMPERATURE overrides the request temperature for OpenAI-compatible backends; none/omit/default omits the parameter. OpenAI reasoning models (o1/o3/o4, gpt-5) always omit temperature automatically.
• claude-cli routes through the locally-installed Claude Code CLI (claude -p --output-format json), so Pro/Max subscribers can run the pass on their plan.

Response cache

A semantic cache stores the extraction result on disk under codegraph-out/cache/semantic, keyed by the corpus content (and per-file by content + relative path). An unchanged corpus on a rebuild reuses the cached result and makes no API call. A cache hit reports zero tokens and zero cost.

Cost awareness

Each backend has a published per-million-token price for its default model, used to estimate the run's USD spend. Local and subscription backends (Ollama, claude-cli) are zero-rated. The end-of-run line ("LLM usage (extraction pass): ... ~$X estimated on ") reports only the dominant extraction-pass tokens; the small tiebreaker and community-labeling prompts are not metered.

Approximate published rates (USD per million input / output tokens, for each backend's default model):

Backend	Input	Output
deepseek	0.14	0.28
openai	0.40	1.60
gemini	0.50	3.00
kimi	0.74	4.66
azure	2.50	10.00
claude	3.00	15.00
bedrock	3.00	15.00
ollama	0.00	0.00
claude-cli	0.00	0.00

These are estimates, not invoices: using a non-default model (via a *_MODEL override) can change the real price, notably for Azure and Bedrock. Because the pass needs a real API key and makes paid calls, it is opt-in and never runs without --semantic.