Search and Research Modes
In Spotter 3, we are introducing Search and Research Mode, which are complementary coexisting modes that share the same underlying toolset while scaling in intensity. Both modes utilize the same agentic tool calls— such as executing Python code, querying the data warehouse, or searching unstructured apps like Slack– but they differ fundamentally in their reasoning and tool budget, and prompt complexity.
Coexistence: speed versus depth
- Search Mode (the sprint)
-
Designed for high-frequency, low-latency insights. It uses a concise “thinking budget” to quickly translate your natural language query into validated insights. It’s your go-to for questions like “What was the revenue in Q3?” where speed and a single, accurate tool call are the priorities.
- Research Mode (the marathon)
-
Built for high-stakes, multi-layered investigations. It takes the same tool-calling capabilities but wraps them in a deeper “reasoning loop”. Instead of a single pass, it consumes a larger computational budget to perform a chain of thought– iteratively querying, validating, and cross-referencing– until it can solve broad prompts like “Diagnose the root cause of our Q3 revenue miss and suggest three recovery paths.”
It might help to think of it like this: If Search Mode is a high-speed calculator that gives you the right sum instantly, Research Mode is the mathematician who shows the entire proof, explores alternative theories, and double-checks for errors before speaking.
Research Mode
Research Mode transforms the standard BI workflow into a proactive, iterative loop designed for “deep analytics,” where Spotter reasons through a problem like a human analyst.
Extended planning (chain of thought): Unlike a single-pass search, Research Mode creates a multi-step query plan to decompose complex questions. It identifies ambiguities, generates testable hypotheses, and allows users to modify the analytical path before execution. Iterative self-correction: If an anomaly is detected, the agent autonomously deep-dives, self-corrects its logic, and refines subqueries until it reaches a grounded conclusion. Automated narrative synthesis: The final output is a coherent report that explains the “why” and “what’s next,” citing specific data points and visualizations to build trust, structured in a detailed report.
Key capabilities
| Feature | Search Mode | Research Mode |
|---|---|---|
Tool Calls |
SQL, Python, App Search (Instant) |
SQL, Python, App Search (Iterative) |
Reasoning Budget |
Minimal: optimized for speed |
Extended: optimized for thoroughness |
Prompt Depth |
Single-dimensional (“What”) |
Multi-dimensional (“Why” and “what if”) |
Output type |
Direct answer or visual |
Comprehensive decision report |
