1# Research Agent
2 
3## Purpose
4 
5The Research Agent gathers, synthesizes, and summarizes information from multiple sources to answer complex research questions. It implements a multi-step research workflow with source verification and citation tracking.
6 
7## Agent Definition
8 
9```typescript
10import { ToolLoopAgent } from "ai";
11import { openai } from "@ai-sdk/openai";
12import { researchTools } from "../tools";
13 
14export const researchAgent = new ToolLoopAgent({
15  name: "researcher",
16  model: openai("gpt-4o"),
17  instructions: `You are an expert research analyst.
18 
19Your role is to:
201. Break down complex research questions into searchable queries
212. Gather information from multiple sources
223. Verify and cross-reference claims
234. Synthesize findings into coherent summaries
245. Provide proper citations for all claims
25 
26Research Methodology:
27- Start with broad searches to understand the landscape
28- Narrow down to specific sources for detailed information
29- Always verify facts from multiple sources when possible
30- Distinguish between facts, claims, and opinions
31- Note the recency and authority of sources
32 
33Quality Standards:
34- Never fabricate information or sources
35- Clearly indicate when information is uncertain
36- Provide direct quotes when precision matters
37- Include source URLs/references for verification`,
38  
39  tools: {
40    webSearch: researchTools.webSearch,
41    readUrl: researchTools.readUrl,
42    extractClaims: researchTools.extractClaims,
43    verifyClaim: researchTools.verifyClaim,
44    synthesize: researchTools.synthesize
45  }
46});
47```
48 
49## Capabilities
50 
51### Web Search
52Search the web for relevant information.
53 
54**Input:**
55- Search query
56- Optional filters (date, source type)
57 
58**Output:**
59- List of relevant results
60- Snippets and URLs
61- Source metadata
62 
63### URL Reading
64Extract content from a specific URL.
65 
66**Input:**
67- URL to read
68- Content type (article, paper, documentation)
69 
70**Output:**
71- Extracted text content
72- Key sections identified
73- Publication metadata
74 
75### Claim Extraction
76Identify distinct claims from a source.
77 
78**Input:**
79- Source text
80- Claim types to extract
81 
82**Output:**
83- List of claims
84- Confidence level
85- Supporting context
86 
87### Claim Verification
88Cross-reference a claim against other sources.
89 
90**Input:**
91- Claim to verify
92- Original source
93 
94**Output:**
95- Verification status
96- Supporting/contradicting sources
97- Confidence assessment
98 
99### Synthesis
100Combine findings into a coherent summary.
101 
102**Input:**
103- Research findings
104- Target format
105- Key questions to answer
106 
107**Output:**
108- Synthesized summary
109- Key insights
110- Source citations
111 
112## Configuration
113 
114```typescript
115interface ResearchConfig {
116  // Search configuration
117  maxSearchResults: number;
118  preferredSources: string[];
119  excludedDomains: string[];
120  
121  // Verification settings
122  minSourcesForVerification: number;
123  requireRecentSources: boolean;
124  maxSourceAge: "1month" | "6months" | "1year" | "any";
125  
126  // Output configuration
127  citationStyle: "inline" | "footnote" | "endnote";
128  summaryLength: "brief" | "standard" | "comprehensive";
129  includeSourceQuality: boolean;
130}
131 
132const defaultConfig: ResearchConfig = {
133  maxSearchResults: 10,
134  preferredSources: [],
135  excludedDomains: [],
136  minSourcesForVerification: 2,
137  requireRecentSources: false,
138  maxSourceAge: "any",
139  citationStyle: "inline",
140  summaryLength: "standard",
141  includeSourceQuality: true
142};
143```
144 
145## Usage Example
146 
147```typescript
148import { researchAgent } from "./agents/research-agent";
149 
150const research = await researchAgent.generate({
151  prompt: `Research the current state of LLM evaluation methods.
152 
153I need to understand:
1541. What are the main approaches to evaluating LLM outputs?
1552. What are the limitations of human evaluation?
1563. How reliable are LLM-based evaluators compared to humans?
1574. What are best practices for implementing LLM-as-a-Judge?
158 
159Provide a comprehensive summary with citations.`
160});
161```
162 
163## Research Workflow
164 
165```mermaid
166graph TD
167    A[Research Question] --> B[Query Decomposition]
168    B --> C[Initial Search]
169    C --> D[Source Selection]
170    D --> E[Deep Reading]
171    E --> F[Claim Extraction]
172    F --> G[Cross-Verification]
173    G --> H[Synthesis]
174    H --> I[Final Report]
175```
176 
177## Integration Points
178 
179- **Knowledge Base Building**: Populate internal knowledge stores
180- **Fact Checking**: Verify claims in generated content
181- **Market Research**: Gather competitive intelligence
182- **Technical Documentation**: Research implementation approaches
183 
184