DeepSeeker — Lightweight Multi-LLM Deep Research Engine

DeepSeeker is a modular deep-research system inspired by OpenAI’s Deep Research, designed for transparent reasoning, iterative search, and multi-model collaboration.

It integrates:

• LLM0 — Planner & Final Analyst (high-quality model)
• LLM1 — Reader & Summarizer (cost-efficient model)
• BingSift — Web search + result extraction backend

DeepSeeker runs a full research pipeline:

1. LLM0 analyzes the question and decides whether to perform web search.
2. If search is required, DeepSeeker uses BingSift to fetch SERP results.
3. LLM0 inspects results (title + snippet) and chooses which pages require deep reading.
4. LLM1 reads each selected page and produces structured summaries.
5. LLM0 synthesizes all summaries and generates a final, well-structured report.

All LLM responses follow a lightweight JSON protocol (MCP-like, but much simpler), ensuring predictable, controllable system behavior.

---

Features

• Two-LLM architecture for optimal cost/performance.
• Deterministic JSON protocol for plan/selection/summarization/synthesis stages.
• Extensible orchestrator written in Python.
• Search powered by BingSift (keyword filtering, domain control, freshness filters).
• Human-readable step logging for full transparency.
• CLI utilities for testing each stage:
  • plan — LLM0 planning behaviour
  • search — BingSift integration
  • run — full research pipeline

---

Installation

pip install -r requirements.txt

Set environment variables:

export OPENAI_API_KEY="your-key"
# Optional: custom endpoint
# export OPENAI_BASE_URL="https://your-host/v1"

Configuration

DeepSeeker supports configuration via JSON file or environment variables.

Option 1: Configuration File (Recommended)

Create a default configuration file:

python -m deepseeker.cli init

This creates config.json in your current directory:

{
  "api_key": "your-openai-api-key",
  "base_url": "https://api.openai.com/v1",
  "llm0": {
    "model": "gpt-5.1-thinking",
    "max_output_tokens": 4096
  },
  "llm1": {
    "model": "gpt-4o-mini",
    "max_output_tokens": 1536
  },
  "search_max_results": 10,
  "search_freshness": "week"
}

Automatic Detection: If config.json exists in the current directory, it will be automatically used by all commands.

Custom Config File: You can also specify a custom config file:

python -m deepseeker.cli --config custom_config.json run --question "your question"

Option 2: Environment Variables

If no config.json file is found, DeepSeeker falls back to environment variables:

# API Configuration
export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://api.openai.com/v1"  # Optional

# LLM Configuration
export DEEPSEEKER_LLM0_MODEL="gpt-5.1-thinking"
export DEEPSEEKER_LLM0_MAX_TOKENS="4096"
export DEEPSEEKER_LLM1_MODEL="gpt-4o-mini"
export DEEPSEEKER_LLM1_MAX_TOKENS="1536"

# Search Configuration
export DEEPSEEKER_SEARCH_MAX_RESULTS="10"
export DEEPSEEKER_SEARCH_FRESHNESS="week"

Note: To use environment variables, either delete config.json or specify a different config file with --config.

Priority Order

Configuration is loaded in this priority order:

1. Explicit --config parameter
2. Default config.json file (if exists)
3. Environment variables
4. Built-in defaults

Configuration Example

Here's a complete example of a config.json file:

{
  "api_key": "sk-your-openai-api-key-here",
  "base_url": "https://api.openai.com/v1",
  "llm0": {
    "model": "gpt-4o",
    "max_output_tokens": 4096
  },
  "llm1": {
    "model": "gpt-4o-mini",
    "max_output_tokens": 1536
  },
  "search_max_results": 15,
  "search_freshness": "week"
}

Tips:

• Leave api_key empty ("") to use environment variable OPENAI_API_KEY
• Leave base_url empty ("") to use default OpenAI endpoint
• You can use different config files for different projects:

  deepseeker --config project1.json run --question "..."
  deepseeker --config project2.json run --question "..."

---

CLI Usage

1. Initialize Configuration

# Create default config.json
python -m deepseeker.cli init

# Create custom config file
python -m deepseeker.cli init --output my_config.json

2. Search

# Uses config.json or environment variables
python -m deepseeker.cli search --query "intel earnings"

# Override config settings
python -m deepseeker.cli search --query "intel earnings" --when week --max-results 20

3. Planning

# Uses config.json or environment variables
python -m deepseeker.cli plan --question "Explain ARM vs RISC-V for servers."

# Use custom config file
python -m deepseeker.cli --config custom.json plan --question "your question"

4. Run full pipeline

# Uses config.json or environment variables
python -m deepseeker.cli run --question "Latest advances in large-scale model training using distributed computing and GPU clusters?"

# Use custom config file
python -m deepseeker.cli --config custom.json run --question "your question"

You will see:

• Markdown final answer
• Key summary points
• A full JSON log of internal steps

Lightweight JSON Protocol

DeepSeeker enforces strict JSON outputs:

LLM0 Plan

{
  "action": "direct_answer" | "search_then_answer",
  "direct_answer": "...",
  "search": {
    "query": "...",
    "when": "week",
    "include": [],
    "exclude": [],
    "allow_domains": [],
    "deny_domains": [],
    "max_results": 10
  },
  "notes": "..."
}

LLM1 Summary

{
  "title": "...",
  "summary": "...",
  "key_points": ["..."],
  "relevance_score": 0.0,
  "notes": "..."
}

LLM0 Final Synthesis

{
  "answer": "markdown report",
  "key_points": [],
  "used_results": [],
  "notes": "..."
}

Roadmap

• Two-LLM orchestration
• BingSift search integration
• Enhanced Log System
• WebUI
• Deep Iterative Search
• Smart Page Reader
• Plugin System

License

MIT License (or replace with your preferred license).