Deep Research and Deep Thinking

Chipp apps support two advanced reasoning capabilities that go beyond standard chat responses: Deep Research for comprehensive multi-source investigation, and Deep Thinking for visible step-by-step reasoning. Both give your consumers more thorough, transparent answers to complex questions.

Chipp uses a two-tier model architecture for Deep Research: a fast, cost-efficient specialist model extracts and normalizes evidence from web sources, while your app’s primary model synthesizes that evidence into a final answer. This separation lets you optimize for both speed and quality without paying for high-end reasoning on every extraction task.

Two-Tier Deep Research Architecture

Deep Research uses two separate models working together:

1. Specialist Model (extraction & evidence normalization): Processes raw web sources and transforms them into a structured, citation-preserving evidence packet. This role uses a cost-optimized model that specializes in efficient source extraction.
2. Main Model (synthesis): Your app’s primary model, which reads the evidence packet and synthesizes a comprehensive, cited answer. This is the same model that powers all other conversations.

This two-tier approach gives you the best of both worlds: you pay for economical source processing while reserving your main model for final synthesis and complex reasoning.

Default Models

New apps ship with Chipp’s recommended defaults:

• Main/Synthesis Model: Grok 4.3 — combines high-quality reasoning with competitive pricing, ideal for synthesis tasks.
• Deep Research Specialist Model: DeepSeek V4 Flash — ultra-fast and cost-efficient evidence extraction.

These defaults are optimized for the cost-sensitive research use case. Most builders never need to change them.

Customizing Models (Advanced)

You can override both models in your app’s Build settings, under Model Selection:

• Main Model: Choose any supported chat model (the same selector you use for general chat). This model synthesizes research findings and powers all other conversations.
• Deep Research Specialist Model: Choose from a curated list of extraction-optimized models. This model only runs on the evidence-extraction step and does not affect synthesis quality.

Model changes apply immediately to new research requests.

Deep Research

Deep Research spawns multiple parallel research agents that independently search the web, read pages, and compile findings. The results are synthesized into a single, comprehensive report with citations.

Invocation Policy

You can control when Deep Research is triggered in your app’s Build settings, under Deep Research:

• Auto (default): The AI uses its own judgment. It activates Deep Research for questions that genuinely benefit from multi-source research (market analysis, comparisons, current events) and skips it for simple questions or conversations referencing existing context.
• Required: The AI must use Deep Research for every research-eligible question. Useful if you want to guarantee comprehensive sourcing for all user queries.
• Off: Deep Research is disabled. The app still answers questions, but without the parallel research capability.

Most apps use Auto and never need to adjust this setting.

When the AI Uses Deep Research

By default (Auto mode), the AI decides to use Deep Research for questions that require broad investigation:

• Market research and competitive analysis
• Technical deep dives requiring multiple authoritative sources
• “Compare X vs Y vs Z” questions with multiple dimensions
• Industry reports and comprehensive analyses
• Multi-faceted questions where a single web search would be insufficient

It does not use Deep Research for simple factual questions, casual conversation, follow-up questions where context already exists, or questions about the consumer’s own data.

How It Works

Citation Integrity

Chipp’s two-tier architecture includes built-in citation integrity:

1. The specialist model extracts raw findings from web sources and produces a structured evidence packet (claims, excerpts, source URLs, publication dates, relevance scores).
2. The specialist’s citations are validated — any claim citing a source that wasn’t in the original web search is dropped.
3. The main model synthesizes the final answer using ONLY the validated evidence packet. It cannot invent new citations; every claim in the final answer links back to the evidence the specialist extracted.
4. If the specialist produces malformed output, the system falls back to the original web sources, ensuring the answer is always grounded and citable.

This means every citation in a Deep Research result is verifiable from the actual sources searched, not inferred or synthesized by the models.

What the Consumer Sees

During Deep Research, the chat UI shows real-time progress:

1. Plan — The sub-questions being investigated
2. Research progress — Which agents are running, how many sources have been found
3. Gap analysis — Whether follow-up research is needed
4. Final report — A comprehensive response with inline source citations

A typical Deep Research operation takes 30-90 seconds depending on the complexity of the question and the number of follow-up investigations needed.

Example Prompts

• “Research the current state of AI regulation in the EU, US, and China”
• “Compare the top 5 CRM platforms for small businesses under 50 employees”
• “What are the latest developments in solid-state battery technology?”
• “Analyze the competitive landscape for AI code assistants”

Deep Thinking

Deep Thinking enables visible, structured reasoning before the AI answers. Instead of jumping directly to a response, the AI creates a step-by-step plan and works through it, showing its reasoning process to the consumer.

Deep Thinking always uses your app’s primary model (the same model that powers all your other conversations). There is no separate specialist model for reasoning — the full depth of your main model is reserved for this task.

Invocation Policy

You can control when Deep Thinking is triggered in your app’s Build settings, under Deep Thinking:

• Auto (default): The AI uses its own judgment. It activates Deep Thinking for multi-part questions and complex analysis tasks where step-by-step reasoning provides clear value.
• Required: The AI must use Deep Thinking for every applicable question. Useful if you want to guarantee transparent, structured reasoning for all analyses.
• Off: Deep Thinking is disabled. The app still answers questions, but without showing step-by-step reasoning.

Most apps use Auto.

When the AI Uses Deep Thinking

The AI automatically uses Deep Thinking for questions that benefit from structured analysis:

• Comparing two or more options, approaches, or strategies
• Questions with multiple distinct parts that each need analysis
• Providing recommendations based on specific criteria
• Complex analysis requiring structured reasoning
• Cross-referencing multiple knowledge base sources
• Combining web research with knowledge base data

It does not use Deep Thinking for simple factual questions with a single clear answer, casual conversation, brief follow-ups, or yes/no questions.

How It Works

What the Consumer Sees

During Deep Thinking, the chat UI shows the reasoning steps as the AI works through them. Each step has a label (e.g., “Searching knowledge base for pricing data”, “Comparing feature sets”, “Synthesizing recommendation”) and updates in real time as the AI progresses.

This transparency helps consumers understand how the AI arrived at its answer and builds trust in complex recommendations.

Example Prompts

• “Should I use PostgreSQL or MongoDB for my e-commerce app?”
• “Review my business plan and identify the three biggest risks”
• “What’s the best marketing strategy for a B2B SaaS launch?”
• “Analyze our Q3 sales data and recommend focus areas for Q4”

Deep Research vs. Deep Thinking

	Deep Research	Deep Thinking
Purpose	Gather and synthesize information from many external sources	Structure reasoning about available information
Speed	30-90 seconds	10-30 seconds
Sources	Parallel web search agents → evidence extraction → synthesis	Knowledge base, web, code execution
Output	Cited research report (citations guaranteed verifiable)	Structured analysis with visible reasoning
Best for	”What is the current state of X?"	"What should I do about X?”
Models	Specialist (extraction) + Main (synthesis)	Main model only
Agents	2-6 parallel sub-agents + follow-up agents	Single agent with multi-step plan

In practice, the AI often combines both: using Deep Research to gather information and Deep Thinking to analyze it. The AI decides which capability to use (or both) based on the nature of the question.

Tier Requirements

Both capabilities require a Builder plan or higher:

Tier	Deep Thinking	Deep Research
Free	No	No
Builder	Yes	Yes
Studio	Yes	Yes
Studio Pro	Yes	Yes
Enterprise	Yes	Yes

Billing and Cost Optimization

Both Deep Research and Deep Thinking consume tokens through your organization’s Stripe Token Billing balance.

Deep Research Costs

Deep Research runs multiple parallel specialist extraction calls (one per sub-question, plus gap analysis) followed by a final synthesis on your main model. Token usage breaks down across stages:

• Specialist stages (extraction, gap analysis): Run on your Deep Research specialist model (default: DeepSeek V4 Flash) — optimized for cost efficiency.
• Synthesis stage: Runs on your main model (default: Grok 4.3) — once, after all extraction is complete.

A typical Deep Research operation uses 5-15x more tokens than a standard chat response. However, because the cost-heavy synthesis step runs only once with validated evidence (not once per sub-question), using a budget specialist for extraction dramatically reduces total cost compared to running every stage on a high-end model.

If you’re doing high-volume research and cost is a concern, the default models are already optimized for that use case. Switching to more expensive models will increase your bill proportionally.

Deep Thinking Costs

Deep Thinking has modest overhead since it uses a single model with structured planning — typically 2-3x a standard response, all on your main model.

Token Transparency

Your organization’s performance dashboard and usage reports break down Deep Research and Deep Thinking token consumption by stage and model, so you can see exactly which models are consuming tokens and optimize your configuration if needed.

Verifying When Capabilities Run

You can see when Deep Research and Deep Thinking were used in your app’s performance dashboard:

• Deep Research triggered: Look at the research progress UI displayed to consumers (plan, research progress, gap analysis). This visibility is the clearest signal.
• Invocation policy: Check your app’s Build settings under Deep Research or Deep Thinking to see whether you’re in Auto, Required, or Off mode.
• Stage-level cost breakdown: Your organization’s usage reports show per-stage token consumption. If you see tokens logged against “specialist-extraction”, “specialist-gap-analysis”, or “main-synthesis”, Deep Research ran. If you see “planning” or “reasoning” stages, Deep Thinking ran.

This transparency helps you understand your token consumption and debug whether capabilities are activating as expected.