Research chains multiple retrieval steps to explore a topic, organize findings, and synthesize a cohesive report. It builds on retrievers, pipelines, and interactions rather than a single API call.
Overview
Multi‑step Search
Break complex questions into sequential, targeted retrieval steps
Structured Synthesis
Organize findings into sections and produce a coherent narrative
Source‑grounded
Maintain citations and lineage back to original documents
Configurable Flow
Compose with retrievers, pipelines, and enrichment stages
How it works
1
Decompose
Split the question into sub‑topics and define per‑topic retrieval intents
2
Retrieve
Execute retrievers per sub‑topic; apply filters, grouping, and selection
3
Enrich
Optionally apply taxonomies or clustering to label and group
4
Synthesize
Assemble a structured report with sections and citations
Example patterns
- Query recent papers, group by venue or year
- Extract abstracts, key findings, and limitations
- Summarize trends with citations back to source docs
Building blocks
Retrievers
Stage KNN, hybrid, filters, grouping, and selection
Pipelines & Workflows
Chain collections and stages to orchestrate multi‑step flows
Interactions
Capture feedback signals to inform reranking across steps
Tasks
Track long‑running enrichment or synthesis jobs
Tips
1
Define scope
Set clear sub‑topics and stop criteria to bound exploration
2
Curate signals
Prefer high‑precision filters and groupings for each step
3
Cite sources
Preserve document ids and metadata to ground conclusions
4
Iterate
Refine prompts, filters, and stage parameters based on gaps