The Design Problem
Chat already had web search. Users wanted more — academic paper analysis, competitive intelligence, autonomous multi-step research, and the ability to build custom research workflows. Bolting features onto the existing chat interface would have created a mess.
We needed a layered system where capabilities compose. Search is always available. Deep Analysis adds multi-phase reasoning on top. Discovery adds product and competitive intelligence. Autonomous Agent runs a full ReAct loop that calls any of the above.
Four Layers, One Workspace
Layer 0: Search
The foundation. 18 search providers — arXiv, Semantic Scholar, PubMed, OpenAlex, Crossref, Tavily, Brave, Product Hunt, GitHub, Kaggle, HuggingFace, and others. Each provider has a normalized response schema, so the layers above don't care where a result came from.
16 focus modes configure the provider mix. Academic mode emphasizes arXiv and Semantic Scholar. Finance mode pulls market data. Space mode queries NASA and ESA databases. Each focus mode defines which providers to query, in what order, and how to merge results.
Four depth levels control thoroughness. Quick runs a single provider. Standard runs the top 3. Deep runs all relevant providers and deduplicates. Pro Search decomposes your query into sub-queries and runs them in parallel across all providers.
Layer 1: Deep Analysis
Multi-phase research methodology built on top of search results.
Phase 1 — Evidence Grading. Every source gets a credibility score based on publication venue, citation count, recency, methodology description, and cross-reference consistency.
Phase 2 — Hypothesis Generation. The system generates competing hypotheses from the evidence, explicitly noting which sources support or contradict each hypothesis.
Phase 3 — Peer Review Simulation. Each hypothesis is challenged — the system plays devil's advocate, identifies methodological weaknesses, and flags areas where evidence is thin.
Phase 4 — Convergence Analysis. Synthesize the surviving hypotheses into a coherent answer with explicit confidence levels and identified knowledge gaps.
A thought tree tracks non-linear reasoning. Branches represent alternative interpretations. You can expand any branch, and the tree maintains causal links between evidence and conclusions.
Layer 2: Discovery
Product and competitive intelligence. Catalog extraction identifies products, features, and pricing from web sources. Competitor landscape mapping visualizes market positioning. Feature matrix comparison generates structured comparisons across dimensions you define.
SWOT analysis and radar charts are generated from the gathered intelligence. Industry presets (SaaS, e-commerce, fintech, healthtech, etc.) configure the discovery tools with domain-specific heuristics.
Layer 3: Autonomous Agent
A ReAct (Reason + Act) loop that plans, executes, and adapts.
The agent has a three-panel monitor:
- Plan — Shows the current strategy and task breakdown
- Feed — Live execution stream showing every tool call, every search, every analysis step
- Evidence — Sources as they're discovered, organized by relevance
You can steer the agent mid-execution — adjust the plan, add constraints, or redirect focus. Token budgets prevent runaway costs. Checkpoints let you save and resume long research sessions. And the agent asks for plan approval before executing, so you're always in control.
29 Research Tools
The tool system is the connective tissue. Every layer above uses the same tools, and tools compose into pipelines.
Search & Discovery (5): Web search, academic search, patent search, dataset search, pro search (query decomposition + parallel multi-provider).
Evidence Analysis (6): Analyze sources, extract structured data, check consistency across sources, detect contradictions, compare results, identify knowledge gaps.
Validation & Quality (6): Assess source credibility, compute confidence intervals, cross-validate claims against multiple sources, assess bias, evaluate methodology fitness, cluster sources by perspective.
Writing & Synthesis (6): Generate hypotheses, evaluate research quality, write report sections, synthesize findings, compute statistical summaries, build event timelines.
Citation & Graph (3): Search citations, build citation graphs with forward and backward traversal, search within document collections.
Product Intelligence (2): Discover products from web mentions, run competitive analysis frameworks.
Each tool has a typed input schema and a typed output schema. Tools can call other tools. The pipeline builder (described below) exposes these same tools as visual nodes.
Visual Pipeline Builder
For repeatable research workflows, the pipeline builder lets you construct a directed acyclic graph (DAG) of research steps.
46 node types span search, analysis, validation, generation, citation, and product categories. Conditional branching supports compound AND/OR logic — "if credibility > 0.8 AND source is peer-reviewed, route to deep analysis; otherwise, flag for manual review."
Independent branches execute in parallel. Auto-layout keeps the graph readable as it grows. Node grouping collapses related steps into labeled subgraphs. Execution visualization shows data flowing through the graph with glow and pulse effects.
Five built-in templates cover common workflows: literature review, competitive analysis, trend research, fact-checking, and systematic review. Version history with snapshot and revert. A/B pipeline comparison runs two workflow variants on the same input and compares results side by side.
The pipeline builder and the Research Builder (a linear step-based alternative) convert between each other automatically. Build a chain in the simple view, switch to the DAG editor for branching, switch back — nothing is lost.
Chat Integration
Research results flow into chat. Insert a finding, a citation, or an entire analysis with one click. The finding carries its provenance — which sources, which tools, which confidence level.
Going the other direction, chat messages escalate to research. Select a message thread, send it to Deep Analysis or Autonomous mode, and the research context picks up where the conversation left off.
A unified activity feed on the Chats page merges both histories, so you can see the full arc of how a question evolved from casual conversation to structured research.
What's Next
The Research Lab is a foundation. The tool system is extensible — we're building toward user-defined tools, community-shared pipeline templates, and research collaboration across teams. The next post in this series will cover the distributed execution architecture that makes parallel pipeline branches possible.