Analyze Module

LLM-based analysis stage for ranked candidates.

This module provides a minimal async analysis function that: - Reduces context (abstract only for now) - Calls the shared LLM model once per paper - Returns structured AnalysisResult instances

Retries and concurrency limits can be layered in the orchestrator later.

async agent.pipeline.analyze.analyze_candidates(*, task_query, analysis_inputs)[source]

Analyze candidates via agents or a heuristic fallback.

When an API key and the environment flag PIPELINE_USE_AGENTS_ANALYZE is set, uses the configured LLM agent to produce structured outputs. Otherwise, computes a quick overlap-based heuristic.

Parameters:

  • task_query (str) – The task description that guides relevance.

  • analysis_inputs (List[AnalysisInput]) – Ranked inputs containing candidates and optional snippets.

Return type:

List[AnalysisResult]

Returns:

One AnalysisResult per input, preserving order.