hidden tulip
Hi, everyone!
I'm building an AI agent in Langgraph for handling complex user queries that require high precision, such as extracting and analyzing multiple parameters from documents (e.g., 20+ items from a PDF like values, estimated timelines, stages, categories, opinions, etc.). The queries often have large inputs with multiple subtópics, so I'm planning this decomposition-based approach to ensure accurate and thorough results. I'd love some feedback on whether this seems solid, or if there are pitfalls/optimizations I'm missing—especially around token costs, latency, and implementation in Langgraph.
Proposed Workflow:
Initial Check (Agent Node): The agent receives the full user input (e.g., a list of 20 extraction parameters from an attached PDF). It calls a main tool that first checks if it's doc-related (if not, respond like a standard GPT). If yes, pass the entire input to a decomposer LLM.
Decomposition: Use GPT-5 or GPT-5-mini (for cost efficiency) to break the input into distinct "missions" (subtasks). Prompt example: "Identify and list distinct subtópics as JSON missions, grouping similar ones to minimize count." This might reduce 20 items to ~10-15 missions by clustering (e.g., combine related details like types and locations).
Parallel Processing of Missions: For each mission, run a "deep research" loop in parallel (using async/parallel edges in Langgraph to cut latency):
Rewrite the mission as a standalone query.
If doc-related: Use RAG tool to extract info.
If needed (e.g., for external references): Fall back to web search.
Self-reflection: After initial result, use the same LLM to check "Is this complete/accurate?" and iterate (max 2-3 times to avoid infinite loops).
Store short summaries + evidence per mission.
Continue below 👇