Deep Knowledge

The Problem

Current summarization tools in the market typically generate shallow, surface-level summaries that fail to capture the rich complexity of content. These summaries often:

• Miss the deeper connections between concepts
• Ignore the hierarchical structure of information
• Extract only the most obvious points
• Lack contextual understanding
• Fail to identify underlying frameworks or mental models

As a result, these tools provide little value for comprehensive learning, critical analysis, or deep content exploration.

Our Approach: Multi-Agent Summarization

Deep Knowledge introduces a powerful multi-agent system for creating deep, comprehensive summaries of complex content. Instead of treating summarization as a single task, we break it down into a coordinated pipeline of specialized agents:

1. Mind Map Agent: Analyzes the structure and concepts of the content, creating both a structural and conceptual map
2. Summary Architect: Designs a modular summary structure with specific instructions for each component
3. Content Synthesizer: Generates each module following the architect's specifications

This approach ensures that summaries retain the original content's structure while revealing deeper patterns, frameworks, and connections.

Key Features

• Deep structured summarization: Creates summaries that capture both structure and conceptual depth
• Smart content handling: Automatically processes various document types with OCR detection
• Token-aware processing: Intelligently manages content to work within model context limitations
• Langchain integration: Seamlessly works with Langchain chat models
• Visual mind mapping: Generates comprehensive mind maps to visualize content structure

Installation

pip install deep-knowledge

Usage Example

from deep_knowledge.summary import Summary

# From a file path
summary = Summary(input_path="my_book.pdf")
summary.run()
print(summary.output)

# From text content
text_content = "..."
summary = Summary(input_content=text_content)
summary.run()
print(summary.output)

Streamlit Demo

We created a Streamlit demo to showcase the Deep Knowledge summarization pipeline. To run the demo, follow these steps:

export PYTHONPATH=$PYTHONPATH:"$(pwd)"
streamlit run demo/streamlit_app.py

Langchain Integration

Deep Knowledge integrates smoothly with Langchain chat models. You can provide your own Langchain chat model, or use the "auto" option which intelligently selects the best available model:

from langchain_openai import ChatOpenAI
from deep_knowledge.summary import Summary

# Using a specific Langchain model
llm = ChatOpenAI(model_name="gpt-4o", temperature=0)
summary = Summary(llm=llm, input_path="article.pdf")

# Using auto mode (automatically selects from available API keys)
summary = Summary(llm="auto", input_path="article.pdf")

In "auto" mode, the system prioritizes:

1. Google Gemini models (if GOOGLE_API_KEY is available)
2. OpenAI models (if OPENAI_API_KEY is available)

This allows for easy experimentation with different LLM providers.

Flexible Input Options

Deep Knowledge accepts multiple input formats:

# From a file path
summary = Summary(input_path="document.pdf")

# From raw text content
summary = Summary(input_content="Your content here...")

# From Langchain Document objects
from langchain_core.documents import Document
documents = [Document(page_content="Content 1"), Document(page_content="Content 2")]
summary = Summary(input_documents=documents)

The library supports various file formats including PDF, DOCX, TXT, Markdown, and EPUB, with automatic OCR detection for scanned documents.

Roadmap

Token Management Optimization

Current implementation can be expensive as it sends the full content to each module creation step. Future improvements will include:

• Having the Summary Architect return metadata about specific content sections for each module

However, using Google Gemini Flash models allows for free/low-cost experimentation

Interactive Refinement and Follow-ups

The current pipeline operates as a single-pass process, but users often need to refine outputs based on initial results.

• Enabling conversational interactions to shape the summarization process
• Allowing targeted refinement of specific modules without regenerating the entire summary

Adding possibility to configure LLM based on step

Now, there's only one LLM that works for all steps. But it would be a good feature to allow experimentation with different LLMs for different steps. E.g.:

• gpt-4o for the Mind Map Agent
• o1 for the Summary Architect
• gemini-2.0-flash for the Content Synthesizer

Improved Reproducibility

We're working on enhancing the system's reliability when LLMs don't precisely follow the expected output format, including:

• More robust output parsing
• Fallback strategies for format deviations
• Better error handling and recovery

Enhanced OCR Capabilities

Plans to expand OCR options include:

• Supporting additional OCR providers
• Implementing local OCR processing options

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.