AgentPack CLI Reference

The AgentPack CLI provides several commands to manage, validate, and query your context packs.

Core Commands

agentpack pack

Pack documents into an agent-friendly context pack.

agentpack pack <input_dir> --out <output_dir> [OPTIONS]

- <input_dir>: The directory containing your raw documents (.txt, .md, .csv, .pdf, .docx, .pptx, .xlsx, .html). - --out: The output directory where the pack will be generated. - --include: Include only files matching these glob patterns (comma-separated). - -i, --ignore, --exclude: Additional glob patterns to exclude. - --no-gitignore: Don't use .gitignore or .agentpackignore rules for filtering files. - --no-default-patterns: Don't apply built-in ignore patterns (e.g. .git, node_modules). - --include-hidden: Include hidden files and directories. - --remove-empty-lines: Automatically strip blank lines from text and markdown files to save tokens. - --fast: Fast mode — use PyMuPDF for PDFs instead of Docling, and skip HNSW index build. Best for quick iteration on small corpora. - --verbose: Enable detailed debug logging. - --quiet: Suppress all console output except errors.

Deprecated: --fast-pdf is an alias for --fast and will be removed in a future release. Use --fast instead.

Config file (agentpack.toml)

Pack settings can be stored in an agentpack.toml file in <input_dir>. CLI flags take precedence over config values.

[pack]
chunk_max_tokens = 800    # max tokens per chunk (default: 800)
chunk_overlap    = 0.15   # overlap fraction between chunks (default: 0.15)
fast             = false  # use fast mode (default: false)
remove_empty_lines = false
include = []              # glob patterns to include
exclude = ["*.log"]       # glob patterns to exclude

agentpack index

Pre-build FTS and vector indexes for a compiled pack. Running this before the first retrieve avoids paying the index-build cost at query time.

agentpack index <pack_dir>

The command is idempotent — if the pack content hasn't changed since the last build, both indexes are reused without rebuilding.

agentpack audit

Generates an audit report for a context pack, highlighting extraction warnings (e.g., empty files, PDFs with no readable text) and statistics.

agentpack audit <pack_dir>

agentpack retrieve

Retrieves top-k evidence chunks from a pack using hybrid search (SQLite FTS5 + HNSW vector index, fused with RRF).

agentpack retrieve <pack_dir> "<query>" --top-k 5 --mode hybrid

- <query>: The search term. - --top-k: Number of results to return (default: 5). - --mode: Search mode — hybrid (default), vector, or fts. - --source: Filter results to chunks whose source_id contains this string (substring match). - --section: Filter results to chunks whose citation.section contains this string (substring match). - --page: Filter results to chunks from this page number.

agentpack validate

Validates the structural integrity of a context pack, ensuring all expected files and tables exist and are properly referenced in the manifest.

agentpack validate <pack_dir>

Evaluation Commands

agentpack eval

Runs a deterministic evaluation benchmark against your pack using queries.yml and gold_evidence.yml. Calculates Hit@K, MRR, Citation Precision, and Context Token Savings.

agentpack eval ./benchmarks/my_dataset

agentpack gen-eval

Evaluates Generative QA using AgentPack. This acts as a smoke test / LLM judge to verify downstream generation behavior.

agentpack gen-eval ./benchmarks/my_dataset --gen-model gemini-1.5-flash --judge-model gemini-1.5-pro --limit 10

- --gen-model: The LLM used for generation (default: gemini-1.5-flash). - --judge-model: The LLM used for judging the answer (default: gemini-1.5-pro). - --limit: Limit the number of queries for a quick smoke test.

agentpack prep-benchmark

Downloads and slices a public dataset into a benchmark format for testing AgentPack.

agentpack prep-benchmark --dataset financebench --sample-size 10

- --dataset: Dataset to slice (e.g., financebench, tatqa, qasper). - --sample-size: Number of documents to sample.