pycsamt.agents.pipeline_agent#

pycsamt.agents.pipeline_agent#

PipelineAgent — LLM-assisted MT processing pipeline selection and interpretation.

Bridges pycsamt.pipeline with the agent framework:

  • Accepts a natural-language "request" key → LLM recommends preset/steps

  • Builds and runs the Pipeline

  • LLM interprets the PipelineResult as a narrative

Available presets: "basic_qc", "noise_reduction", "full_processing", "tensor_analysis", "dimensionality_filter", "publication_ready".

Classes

PipelineAgent(*[, api_key, model, ...])

LLM-assisted MT processing pipeline selection and interpretation.
class pycsamt.agents.pipeline_agent.PipelineAgent(*, api_key=None, model=None, llm_provider='claude', preset='basic_qc', param_overrides=None)[source]#

Bases: BaseAgent

LLM-assisted MT processing pipeline selection and interpretation.

Two modes of operation:

Guided mode — pass "request" in input_data. The LLM recommends a preset and any parameter overrides, then the pipeline is built and run automatically.

Direct mode — pass "preset" (name string) or "steps" (list of step codes) in input_data. No pre-run LLM call is made; the pipeline is built directly from those instructions.

In both modes the post-run LLM call interprets the PipelineResult as a narrative.

Parameters:

  • api_key (str) – Standard LLM configuration inherited from BaseAgent.

  • model (str) – Standard LLM configuration inherited from BaseAgent.

  • llm_provider (str) – Standard LLM configuration inherited from BaseAgent.

  • preset (str, optional) – Default preset name used when input_data contains neither "preset" nor "steps" nor "request". Defaults to "basic_qc" (safe first-pass).

  • param_overrides (dict, optional) – Default parameter overrides applied on top of any preset or step list. Format: {step_code: {param: value}}.

  • keys (Output data)

  • ----------

  • path (sites /) – Raw MT/AMT sites to process.

  • request (str, optional) – Natural-language description of dataset and goals. Triggers a pre-run LLM call that recommends preset + parameter overrides.

  • preset – Named preset — overrides constructor default.

  • steps (list of str, optional) – Explicit ordered list of step codes. Ignored when "preset" is set.

  • param_overrides – Per-step parameter overrides — merged on top of constructor defaults.

  • output_dir (str or None, optional) – Root directory for pipeline output files (EDI, plots, YAML config).

  • keys

  • ----------------

  • agents (sites_out Processed Sites — ready for downstream)

:param pipeline_result PipelineResult object: :param preset_used Name of the preset that was run (or "custom"): :param steps_run List of step code strings that were executed: :param n_sites_in Number of input stations: :param n_sites_out Number of stations after processing: :param n_errors Number of steps that raised an error: :param recommendation Dict returned by LLM pre-run call (or None):

Examples

Guided mode:

agent  = PipelineAgent()
result = agent.execute({
    "sites": sites,
    "request": "50 Hz grid noise, possible static shift, Occam2D target",
    "output_dir": "willy_pipeline/",
})
processed = result["sites_out"]
print(result.llm_interpretation)

Direct mode:

agent  = PipelineAgent(preset="full_processing")
result = agent.execute({
    "sites": sites,
    "param_overrides": {"NR001": {"mains_hz": 60}},
})

Chain with Occam2DAgent via AgentCoordinator:

from pycsamt.agents import AgentCoordinator, PipelineAgent, Occam2DAgent

coord = AgentCoordinator("willy_full")
coord.add_step("pipeline", PipelineAgent(preset="full_processing"),
               input_fn=lambda r: {"sites": r["load"].data["sites"]})
coord.add_step("invert",   Occam2DAgent(),
               input_fn=lambda r: {"sites": r["pipeline"].data["sites_out"]})
SYSTEM_PROMPT: str = 'You are an expert MT data processing specialist reviewing pipeline execution results.\nGiven a processing summary, write 3–4 sentences that:\n1. Describe the overall data quality change (stations retained, step errors if any).\n2. Highlight which steps were most impactful or took the most time.\n3. Comment on whether the stated processing objectives appear to have been met.\n4. Recommend a concrete follow-up action (further correction, QC plot review,\n   or readiness for inversion).\nReply in plain English. No bullet points or markdown.\n'#

Override in subclasses to give the LLM its domain expertise.

execute(input_data)[source]#

Run this agent on input_data and return an AgentResult.

Subclasses must implement this method. The contract:

  • Reset self._last_cost = 0.0 at the top.

  • Record wall-clock time with t0 = time.time().

  • Return AgentResult(elapsed_seconds=time.time()-t0, cost_estimate_usd=self._last_cost, ...).

Parameters:

input_data (dict[str, Any])

Return type:

AgentResult