# Design a Multi-Agent Document Search and Visualization Product Design a document-analysis product that searches PDF content, extracts figures and tables, produces structured tables and charts, and streams an answer. The proposed workflow has a status agent, an orchestration agent, specialized search agents, and a final streaming agent. ### Constraints & Assumptions - Documents may contain scanned pages, embedded text, figures, and tables. - Every extracted value must retain page-level provenance. - The UI needs progress updates and partial results without presenting unverified values as final. - Agent retries must not duplicate expensive extraction jobs. ### Clarifying Questions to Ask - What user question types and document sizes must be supported? - Which outputs require exact extraction and which allow semantic retrieval? - What latency budget applies to first status, first evidence, and final answer? ### Part 1 — Architecture Define the responsibilities and contracts of orchestration, retrieval, extraction, status, and answer components. Explain whether each component must be an agent. #### What This Part Should Cover - Typed task and evidence schemas - Deterministic tools for OCR, table parsing, and charting - Durable job state and idempotent retries ### Part 2 — Quality and deployment Design evaluation, provenance, human-review thresholds, streaming behavior, and deployment monitoring from frontend through backend. #### What This Part Should Cover - Golden document-question sets - Cell- and citation-level accuracy - Backpressure, cancellation, cost, and failure observability ### What a Strong Answer Covers - A reasoned agent-versus-tool boundary - End-to-end provenance - Safe partial-result and retry semantics ```hint Prefer typed evidence over agent prose Have retrieval and extraction components return structured evidence with document, page, region, method, and confidence. The answer layer should consume that record rather than untraceable text. ``` ### Follow-up Questions - How would you handle a chart whose axes are visible but data labels are not? - How would you prevent one slow document from blocking the whole answer?