Skip to content

An AI-powered search engine with a generative UI

refactor: strengthen todoWrite guidance for complex queries

Yoshiki Miura3890dda63d32
diff --git a/lib/agents/prompts/search-mode-prompts.ts b/lib/agents/prompts/search-mode-prompts.ts
index 5f612fd9..21c5ff7c 100644
--- a/lib/agents/prompts/search-mode-prompts.ts
+++ b/lib/agents/prompts/search-mode-prompts.ts
@@ -62,27 +62,37 @@ Structure your responses with:
- Summary or conclusion when appropriate
`
-export const AUTO_MODE_PROMPT = `
+export const ADAPTIVE_MODE_PROMPT = `
Instructions:
You are a helpful AI assistant with access to real-time web search, content retrieval, task management, and the ability to ask clarifying questions.
APPROACH STRATEGY:
-1. **Assess query complexity first:**
- - Simple queries (1-2 aspects): Direct search and respond
- - Medium queries (3-4 aspects): Consider using todoWrite for organization
- - Complex queries (5+ aspects or requiring deep research): ALWAYS use todoWrite
+1. **FIRST STEP - Assess query complexity:**
+ - Simple queries (1-2 specific questions): Direct search and respond
+ - Medium queries (3-4 related aspects): SHOULD use todoWrite for organization
+ - Complex queries (ANY of the following): MUST use todoWrite
+ * 5+ aspects to research
+ * Requires comparing multiple viewpoints
+ * Needs systematic investigation
+ * Involves both research AND analysis/synthesis
+ * User asks for "comprehensive" or "detailed" analysis
-2. **For queries with multiple aspects or requiring systematic research:**
- - Use todoWrite to break down the query into clear tasks
- - Update task status as you progress through your research
- - This helps users track your progress and ensures thoroughness
+2. **When using todoWrite (for medium/complex queries):**
+ - Create it as your FIRST action to show your plan
+ - Break down into specific, measurable tasks like:
+ * "Search for [specific aspect]"
+ * "Fetch detailed content from top 3 sources"
+ * "Compare perspectives from different sources"
+ * "Synthesize findings into comprehensive answer"
+ - Update task status as you progress (provides transparency)
+ - This ensures thoroughness and helps users track progress
3. **Search and fetch strategy:**
- - Start with type="general" search to get an overview and identify key sources
- - ALWAYS follow up promising search results with fetch tool for deeper analysis
- - Use multiple searches with different keywords for comprehensive coverage
- - Pattern: Search → Identify top sources → Fetch detailed content → Synthesize
+ - Use type="optimized" for research queries (immediate content)
+ - Use type="general" for current events/news (then fetch for content)
+ - Pattern: Search → Identify top sources → Fetch if needed → Synthesize
+ - Multiple searches with different angles for comprehensive coverage
4. **If the query is ambiguous, use ask_question tool for clarification**
diff --git a/lib/agents/researcher.ts b/lib/agents/researcher.ts
index 784ecfdd..f7cd9c02 100644
--- a/lib/agents/researcher.ts
+++ b/lib/agents/researcher.ts
@@ -16,7 +16,7 @@ import { getModel } from '../utils/registry'
import { isTracingEnabled } from '../utils/telemetry'
import {
- AUTO_MODE_PROMPT,
+ ADAPTIVE_MODE_PROMPT,
PLANNING_MODE_PROMPT,
QUICK_MODE_PROMPT
} from './prompts/search-mode-prompts'
@@ -109,7 +109,7 @@ export function researcher({
case 'adaptive':
default:
// Adaptive Mode: Balanced approach, current behavior
- systemPrompt = AUTO_MODE_PROMPT
+ systemPrompt = ADAPTIVE_MODE_PROMPT
activeToolsList = ['search', 'fetch']
if (writer && 'todoWrite' in todoTools) {
activeToolsList.push('todoWrite', 'todoRead')
diff --git a/lib/config/model-types.ts b/lib/config/model-types.ts
index a6ca0beb..b6430b91 100644
--- a/lib/config/model-types.ts
+++ b/lib/config/model-types.ts
@@ -10,7 +10,8 @@ export const MODEL_TYPE_CONFIGS: Record<ModelType, Model> = {
providerId: 'gateway',
providerOptions: {
gateway: {
- order: ['groq', 'deepinfra']
+ order: ['groq', 'deepinfra'],
+ only: ['groq']
}
}
},