AI Research Agent

A professional-grade AI research platform that delivers comprehensive, multi-source synthesized reports with real-time financial data integration.

🚀 Overview

AI Research Agent is designed for high-stakes information gathering. Unlike standard chat AI, it follows a rigorous research methodology: analyzing the knowledge space, clarifying intent, planning the investigation, and synthesizing findings from across the web and financial markets.

🛠️ Tech Stack

• Framework: Next.js 15+ (App Router, Turbopack)
• LLM Orchestration: OpenCode Zen (utilizing the big-pickle reasoning model for synthesis, via its OpenAI-compatible API).
• Web Search: LangSearch for comprehensive web crawling and information retrieval.
• Financial Data: Alpha Vantage for real-time stock quotes and company overviews.
• Runtime: Bun for ultra-fast development and build cycles.

🖥️ Interactive Research Interface

The application features a dynamic, state-aware interface that guides users through the research process:

• Real-time Feedback: Displays specific loading states (Exploring, Clarifying, Confirming, Researching) so users know exactly what the agent is doing.
• Contextual Search Box: The search input adapts to the current state. During the "Refine Scope" phase, it acts as a context-injector rather than a new search trigger.
• History Management: Automatically saves research sessions to local storage for quick reference.
• Source Transparency: Every report includes direct links to the original sources, ensuring all findings are verifiable.

🔄 Detailed Research Workflow

AI Research Agent follows a sophisticated multi-stage process to ensure accuracy and depth.

1. Exploration & Initial Analysis

• Knowledge Mapping: The agent performs an initial exploratory search to map the knowledge space and identify key themes.
• Ambiguity Detection: Analyzes the query for missing context or multiple interpretations.
• Financial Trigger: Automatically detects if the query involves companies or market data requiring Alpha Vantage integration.

2. Intent Alignment (Clarification)

• Targeted Questions: If the query is broad, the agent generates 3-5 specific questions to narrow the focus.
• User Preferences: Users can answer these questions or "Skip" to let the AI use its best judgment based on initial findings.
• Initial Synthesis: Provides a "Research Scope" summary based on the initial exploration.

3. Strategy & Scope Confirmation

After clarification, the agent presents a Research Strategy which includes:

• Confirmed Scope: A concise summary of what will be investigated.
• Optimized Query: The refined search string that will be used for retrieval.
• Investigation Path: A step-by-step plan showing the specific topics to be covered.

🖱️ User Actions at this Stage:

• Execute Research ("Go"):
  • Confirms the current strategy.
  • Triggers the full multi-source investigation.
  • Moves directly to the final synthesis phase.
• Refine Scope:
  • Allows the user to provide additional instructions or context.
  • Logic: The system combines the previous refined query with the new user input (e.g., "Previous Query + Additional context: [User Input]").
  • Restart: The workflow loops back to Phase 1 with the newly enriched query to re-evaluate the strategy.

4. Comprehensive Investigation & Synthesis

• Query Decomposition: The main query is broken down into 5-7 specialized sub-queries for maximum coverage.
• Parallel Retrieval: Executes simultaneous searches across the web using LangSearch.
• Data Augmentation: Merges real-time financial metrics (if applicable) with web findings.
• Professional Synthesis: OpenCode Zen's big-pickle model processes hundreds of data points to generate a structured, cited report.

📊 Report Structure

Every research report is generated with a professional layout:

• Executive Summary: High-level overview of the findings.
• Key Findings: Bulleted list of critical insights.
• Detailed Analysis: Deep dive into specific aspects of the research.
• Financial Overview: (Optional) Real-time stock data and company metrics.
• Sources: Numbered citations linking directly to the original web sources.

⚙️ Environment Setup

Copy .env.example to .env.local (or .env) in the root directory and fill it in:

# Core API Keys
OPENCODE_API_KEY=your_key_here
# Optional — defaults to this URL
OPENCODE_BASE_URL=https://opencode.ai/zen/v1
LANGSEARCH_API_KEY=your_key_here
ALPHA_VANTAGE_API_KEY=your_key_here

# App Configuration
NEXT_PUBLIC_APP_URL=https://your-deployment-url.com

Env files holding real keys are git-ignored — only .env.example is committed.

LLM notes

The app talks to OpenCode Zen's OpenAI-compatible endpoint (POST {OPENCODE_BASE_URL}/chat/completions) with plain fetch.

big-pickle is a reasoning model: it returns a reasoning_content field and those reasoning tokens are charged against max_tokens. Budgets below ~1024 get consumed entirely by reasoning and return empty content, so token budgets live in LLM_MAX_TOKENS (src/lib/constants.ts) — 16384 for synthesis reports, 4096 for the shorter structured-JSON steps. Model limits: 200k context / 160k input / 32k output, text-only (no vision, embeddings or audio).

🏃 Getting Started

1. Install Dependencies:

   bun install

2. Run Development Server:

   bun dev

3. Build for Production:

   bun run build