P4LTL_LLM

This package contains a retrieval-oriented intent-to-P4LTL system.

Main modules

• syntax_checker/ Parser-backed P4LTL syntax checker and agent-facing validation interface.
• context_store.py Loads P4 sources, build artifacts, runtime JSON, and guide content into memory.
• context_tools.py Read-only retrieval tools for code search, JSON queries, bounded snippet reads, and graph queries.
• intent_decomposer.py Structured intent decomposition layer.
• context_validator.py Checks whether referenced fields/tables/actions/keys match the aligned context.
• semantic_reviewer.py Heuristic semantic review for generated specs.
• pipeline_protocol.py End-to-end IntentToP4LTLPipeline orchestration.
• benchmark_specs.py Default benchmark catalog for Case Study and SageFuzz/P4 programs.
• benchmark_runner.py Runs reference benchmark validation or a pipeline over the benchmark catalog.

Quick usage

Run the default reference benchmark validation:

PYTHONPATH=/home/gosh/P4LTL python3.12 /home/gosh/P4LTL/P4LTL_LLM/benchmark_runner.py

Use the pipeline from Python:

from P4LTL_LLM import IntentToP4LTLPipeline, IntentToP4LTLRequest

pipeline = IntentToP4LTLPipeline(use_agents=False)
request = IntentToP4LTLRequest(
    intent="生成一个简单性质，验证 ingress_port 始终为非负",
    p4_program_paths=["/abs/path/to/program.p4"],
    artifact_paths=["/abs/path/to/build.json"],
    guide_path="/home/gosh/P4LTL/P4LTL_LLM/P4LTL_user_guide",
)
result = pipeline.generate_and_validate(request)
print(result.model_dump_json(indent=2))

Run the online Agno pipeline directly:

PYTHONPATH=/home/gosh/P4LTL python3.12 /home/gosh/P4LTL/P4LTL_LLM/agno_online_generate.py \
  --intent "验证 ingress_port 为 1 的 IPv4 包最终会从端口 3 转发" \
  --admin-description "Use exact fields from the program and artifacts." \
  --p4 /abs/path/to/program.p4 \
  --artifact /abs/path/to/build.json

The online model configuration is now read from:

/home/gosh/P4LTL/P4LTL_LLM/config/api_config.json

Notes:

• Online mode uses IntentToP4LTLPipeline(use_agents=True).
• By default it enables streaming model calls because the configured gateway may reject non-streaming requests.
• Pass --no-stream only if your backend supports non-streaming structured responses.