LLM Enrichment

After AST parsing, Vibe Analyzer can enrich results via Ollama: adding a description and search tags to each file, and a brief summary to the project based on the README.

How It Works

AST data → batching → Ollama request → description + tags for each file

1. Project Summarization

The README (if present) is sent to Ollama with the prompt “write 2-3 sentences about the project”. The result is saved in summary.

2. File Enrichment

Batching. Files are grouped into batches based on two config limits:

max_chunk_chars — maximum characters per request (default 4000)
max_chunk_files — maximum files per request (default 3)

The prompt does not contain the files themselves, but their AST: functions, classes, structs, and other elements. Which elements to include is controlled by the flags ast_imports, ast_variables, ast_functions, ast_enums, ast_interfaces.

Prompt. Ollama receives a JSON template:

{
  "files": [
    {
      "path": "src/main.rs",
      "description": "FILL_DESCRIPTION",
      "tags": ["TAG1", "TAG2", "TAG3"]
    }
  ]
}

The model must fill in description and tags while preserving the structure. The prompt strictly requires: don’t skip files, don’t change paths, copy the JSON as-is.

Response. Ollama returns the completed JSON:

{
  "files": [
    {
      "path": "src/main.rs",
      "description": "Main entry point for the CLI application",
      "tags": ["entry-point", "cli", "argument-parsing"]
    }
  ]
}

3. Parallel Processing

If multiple Ollama hosts are configured, files are distributed among them:

All hosts read from a single channel
The fastest one processes the most
If any host errors — all stop (error_flag)
At the end, per-host statistics are reported: how many files each processed

4. JSON Repair

LLMs often corrupt JSON: add comments, wrap in markdown blocks, drop quotes. clean_llm_json fixes this:

Extracts JSON from ``` blocks
Adds missing key quotes
Removes trailing commas
Balances unclosed braces

5. Retries

If Ollama returns fewer files than were in the batch — up to 5 retries with delay. If after 5 attempts it still doesn’t match — an error with recommendations to reduce max_chunk_chars or switch models.

Generation Parameters

The following are passed from config to the Ollama request:

temperature: 0.1 — low temperature for stable results
seed: 42 — fixed seed for reproducibility
num_ctx: 4096 — context window size
num_predict: 2048 — maximum tokens in response
timeout_secs: 60 — request timeout

Model Warm-Up

Before enrichment begins, for each Ollama host:

Availability check (GET /)
Model presence check (GET /api/tags)
Empty request to load the model into memory (POST /api/generate with empty prompt)

Enrichment Result

After processing, each file receives description and tags from the LLM, and the project receives a summary based on the README.

Exporting Results

AST parsing and LLM enrichment results can be exported to a file for analysis, debugging, or use in other tools:

# AST with LLM enrichment
vibe-analyzer scan analyze --target my-app

# With format and path specified
vibe-analyzer scan ast --target my-app --format json5 --output analysis.json5

Keyboard shortcuts

Vibe Analyzer