When to use this section

Please make sure the search agent is successfully installed and running before testing out any of the examples below!

Note

As a reminder, the search agent currently does NOT work on Brave browser.

Multi-Model Support

Agentic FinSearch supports multiple foundation models.

Available Models

Supported Models:

Model ID

Provider

Underlying Model

Description

FinGPT

Google

gemini-3-flash-preview

Default model. 1M token context window. Best for comprehensive analysis.

FinGPT-Light

OpenAI

gpt-5.1-chat-latest

Fast and efficient. 128K token context window.

Buffet-Agent

Custom

Hugging Face endpoint

Fine-tuned financial model with specialized knowledge.

Switching Models

  1. Click the Settings button in the agent popup.

  2. Select your preferred model from the dropdown.

  3. The agent will use the selected model for all subsequent queries.

MCP (Model Context Protocol) Integration

MCP enables advanced agent capabilities through tool integration.

MCP Features

  • Yahoo Finance MCP: Directly fetches real-time market data, stock prices, and company profiles via the yfinance API.

  • SEC-EDGAR MCP: Enables the agent to directly access SEC filings like 10-K, 10-Q, 8-K and extract financial data from them.

  • TradingView MCP: Fetches technical analysis indicators, oscillators, moving averages, and market screener data.

  • Filesystem MCP: Provides read access to local data files within the application directory.

Deep Research Mode

For complex financial questions that require synthesizing information from multiple sources, the agent offers a Deep Research Mode.

How It Works

  1. Query Decomposition: The agent’s QueryAnalyzer breaks your question into typed sub-questions (numerical, qualitative, analytical).

  2. Parallel Execution: The ResearchExecutor routes each sub-question to the best source — MCP tools for numerical data, web search for qualitative context.

  3. Gap Detection: The GapDetector identifies any missing information and triggers follow-up searches.

  4. Synthesis: The Synthesizer combines all findings into a coherent, well-sourced response.

To use deep research mode, click the Advanced Ask button or select “research” mode via the API.

Note

Research mode typically takes 15-90 seconds depending on query complexity. The agent performs multiple parallel searches and synthesizes the results.

Custom URL Preferences

Configure preferred financial websites for the agent to search.

Setting Preferred URLs

  1. Open Settings

  2. Navigate to Preferred Links

  3. Add URLs of trusted financial sources

  4. Save your preferences

The agent will prioritize these sources when using Advanced Ask.

Example preferred URLs:

  • https://www.sec.gov/edgar

  • https://investor.apple.com

  • https://www.federalreserve.gov

Advanced Query Techniques

Query Modes

Basic Ask: - Searches only the current webpage - Faster responses - Best for page-specific questions

Advanced Ask: - Searches the open domain and uses MCP tools - Activates the deep research pipeline for complex queries - More comprehensive responses - Best for research and analysis

Effective Prompting

For best results:

  1. Be Specific: “What was Apple’s Q3 2024 revenue?” vs “Tell me about Apple”

  2. Request Sources: Add “with sources” to get citations

  3. Compare Data: “Compare Tesla’s P/E ratio to industry average”

  4. Time-bound Queries: Include dates for historical data

Monitoring Agent Activity

Real-time Logs

If you are running the agent locally, monitor the agent’s search and scraping activity:

  1. Inside your terminal and / or browser Dev Tool (console), watch for:

    • MCP status

    • URLs being scraped

    • Search queries executed

    • Model API calls

    • Citations if using advanced ask

    • Error messages

Debug Mode

For troubleshooting:

# Set debug environment variable
export DJANGO_DEBUG=True

# Then restart the backend server via docker, uv or python

Performance Optimization

Smart Context Management

Agentic FinSearch includes a two-tier context management system:

  • Unified Context Manager (default): Session-based context tracking with JSON structure for fast, in-memory conversation management.

  • Mem0 Context Manager (optional): Production-grade long-term memory powered by Mem0 for sessions exceeding 100,000 tokens.

The active mode is configured via the CONTEXT_MANAGER_MODE environment variable (unified or mem0).

How it works:

  • Session-Based: The agent maintains the full conversation history for the current session.

  • Smart Compression: When the context exceeds 100,000 tokens, the agent automatically extracts key facts, entities, and research findings into long-term memory.

  • Fact Preservation: Critical financial data, URLs, and research objectives are preserved while redundant boilerplate is discarded.

  • Session Isolation: Each browser tab/session maintains its own isolated context.

Note

Each browser tab maintains its own conversation context. Refreshing the page starts a new session unless a custom session ID is used.

Troubleshooting Advanced Features

Common Issues

MCP features not working:

  • Confirm OpenAI API key is valid

  • Check you’re using an MCP-compatible model

  • Monitor terminal for MCP-related errors. If errors directly from the MCPs exist, contact Felix via Discord or WeChat.

Slow responses with Advanced Ask:

  • Reduce number of preferred URLs

  • Check internet connection

  • Research mode searches multiple sources in parallel and synthesizes results, which typically takes 15-90 seconds