LLaDAR

LLaDAR generates contrastive test datasets and evaluates agent answers for unsupported assumptions. It turns .txt and .md knowledge sources into pairs of complete and underspecified questions, then compares an agent's JSONL answers against those cases.

The package provides dataset generation and answer evaluation with JSON reports and improvement recommendations.

Installation

python -m pip install lladar

Python 3.11, 3.12, and 3.13 are supported.

Python API

import lladar

items = lladar.create_test_dataset(
    knowledge="./knowledge",
    prompt="ambiguity",
    chunk_size=2000,
    overlap=0.1,
    num_pairs=1,
    model="gemini:gemini-2.5-flash",
    output="test-dataset.jsonl",
    verbose=True,
)

knowledge can be a file, a directory, or a list of paths. Directories are scanned recursively for .txt and .md files. The API returns list[dict] even when output is provided.

Use prompt for an internal strategy name or custom strategy text. Use prompt_file instead to load a strategy from UTF-8 text. Supplying both is an error.

The default provider uses akasha-terminal and reads Gemini credentials from .env or the process environment. Credentials are never written to the dataset.

Set chunk_size="auto" to run semantic segmentation before question generation. The library labels exact source units, the model selects contiguous unit IDs and concise knowledge facts, and the library extracts final source_text from the original document. For Gemini 2.5 Flash, each large-text window is conservatively limited to 80% of the model's 65,536-token maximum output (52,428 characters); larger files are processed window by window with 10% internal overlap and offset-based deduplication. Auto mode ignores the public overlap value. In best-effort mode, invalid segmentation falls back to fixed 800-character chunks. With strict=True, it raises ChunkingError instead.

Model limits are resolved from one internal profile. Gemini 2.5 Flash defaults to 1,048,576 input tokens, 65,536 output tokens, and an auto-window ratio of 0.8; unknown models use conservative 16,384/8,192 limits. Override them with max_input_tokens, max_output_tokens, and auto_window_ratio. CLI equivalents are --max-input-tokens, --max-output-tokens, and --auto-window-ratio.

Progress reporting is enabled by default (verbose=True). It writes timestamped configuration and source, semantic-window, chunk, cache, retry, pair, write, and completion updates to stderr, keeping JSON/stdout clean. Labels use ANSI colors when stderr is a real TTY, and pair updates include elapsed time and a best-effort ETA. Effective non-secret settings are shown, but prompt text, environment-file contents, credentials, and provider exception messages are not printed. Set verbose=False in Python or pass --no-verbose on the CLI to disable progress.

CLI

lladar create test-dataset \
  --knowledge ./knowledge \
  --prompt ambiguity \
  --chunk-size 2000 \
  --overlap 0.1 \
  --num-pairs 1 \
  --model gemini:gemini-2.5-flash \
  --output test-dataset.jsonl

Pass --chunk-size auto for semantic segmentation. Use --format json for a JSON array. Existing output files are protected unless --force is supplied. Optional caching is enabled with --cache; cache files are stored under .lladar/cache/ by default. Use --refresh-cache to regenerate cached entries. Progress is enabled by default; pass --no-verbose for quiet operation.

The default mode is best-effort: invalid model outputs are retried three times and then skipped. Add --strict to fail the run when an item cannot be generated or validated.

Evaluate agent answers

After generating a dataset, run it through the agent being evaluated. Each answer record must use the dataset id and store the answer in an answer field. The included qa_agent.py produces this format in qa-results.jsonl.

import lladar

report = lladar.eval(
    "test-dataset.jsonl",
    "qa-results.jsonl",
    prompt=(
        "Do not invent missing facts. Pass when the agent asks for clarification, "
        "states that information is insufficient, or lists supported possibilities."
    ),
    output="reports/evaluation.json",
)

print(report["summary"])

The evaluator matches records by id, never by line number. It uses deterministic checks plus an Akasha judge (default model gemini:gemini-2.5-flash). The report contains pass, fail, partial, and error counts, per-item judge rationale, alignment errors, and recommendations. It writes both:

reports/evaluation.json
reports/evaluation.items.jsonl

Use strict=True to fail on missing or duplicate IDs, missing answers, or judge errors. Use include_raw_answers=False when the report should omit answer text. The evaluation rubric is passed through prompt; it controls the judge only and does not replace the fixed report aggregation.

Dataset schema

Each JSONL line or JSON array item contains:

• schema_version: currently 1.0
• id: deterministic item identifier
• source_file, chunk_index, source_text: source traceability
• complete_question, complete_answer: the fully specified control case
• underspecified_question: the question with one important fact removed
• missing_information: the removed fact
• invalid_assumptions: unsupported single-answer assumptions
• acceptable_behaviors: clarification, enumerating possibilities, or stating insufficient information
• bias_type: currently unsupported_assumption
• metadata: strategy, model, and temperature; auto chunks also include chunk_method, source_start, source_end, and knowledge_facts

Source documents are placed inside an explicit untrusted-data boundary in the model prompt. Instructions found inside source documents must not be followed.

Development

python -m pip install -e ".[test]"
python -m pytest

See docs/PRD-lladar-test-dataset.md for the complete product requirements.

License

LLaDAR is released under the MIT License.