Intro
Subscribe
Try It Yourself!
Ready to transform your documents into dynamic knowledge? Download our cutting-edge Deep Research Knowledge Graph Extraction Protocol today. Experience firsthand how complex narratives are converted into structured, interconnected insights, unlocking powerful new capabilities for your research and applications.

Reports generated by Deep Research make excellent documents for knowledge graph extraction. Knowledge graphs extracted from a report may be used to iteratively improve report quality by feeding Deep Research the KG and a prompt to use search or documents you provide to fill in gaps in the graph.
Adapting the Protocol to Your Documents
Instead of manually tweaking the research protocol yourself, you can have the AI do it for you. The existing protocol contains information that helps Deep Research modify the protocol for other documents. Here's how to adapt the Deep Research protocol using one or more sample documents:
Provide Deep Research with the Protocol Document
  1. On GitHub
  • Click the “Download raw file” icon.
  • When prompted, choose Save As…, pick your target folder, and save the .md file.
  1. In ChatGPT’s Deep Research interface
  • Switch to your Deep Research chat.
  • Drag the saved .md file from your folder directly into the chat input area.
  • Release to upload—the protocol will appear as an attached Markdown document.
Provide Sample Document(s)
In the Deep Research chat interface (e.g. OpenAI ChatGPT or a similar platform), upload or paste the text of one or more documents that are representative of the format or domain you care about. For example, if you want to adapt the protocol for legal contracts, you might paste a sample contract text.

Instruct the AI to Adapt the Protocol
Ask the Deep Research tool (via a prompt) to "look at the provided document(s) and generate a revised Deep Research protocol tailored for this type of document." Be clear about what type of document it is (business report, legal text, etc.).
Receive a Fully Revised Protocol
The AI will analyze the sample content and output a complete, updated set of guidelines (a full protocol) optimized for researching that kind of document. You won't need to interpret any diffs or change logs – the output will be the entire protocol, already adjusted.
Use the Adapted Protocol
Save the new protocol text and upload it with your target document for graph extraction in the Deep Research UI. You can use it for analyzing similar documents, copy and paste it into your knowledge base, or share it with colleagues.
By following these steps, you leverage Deep Research's capabilities to customize its own instructions based on the document type. This ensures the research process is fine-tuned for the nuances of your content.
Common Document Types & How the Protocol Adapts
Deep Research can handle a wide range of document types. Here are some common examples and what to expect when the protocol is adapted for each:
Business Reports
Typically structured with executive summaries, data charts, and conclusions. The adapted protocol will focus on identifying key performance indicators, summarizing findings clearly, and possibly suggesting visual aids for data. It may emphasize concise bullet points for insights and clear headings for sections like Summary, Analysis, and Recommendations.
Legal Texts (Contracts, Policies, etc.)
Often dense with formal language, definitions, and clause references. The protocol for legal documents will emphasize clarity and accuracy—ensuring terms are defined, obligations are clearly outlined, and referencing specific sections or clauses. It might instruct the AI to be cautious with interpretations and to preserve exact wording for legal precision.
Product Specifications/Technical Docs
These include technical requirements, features, and schematics. The adapted protocol will guide the AI to outline technical details methodically, perhaps creating tables or lists of features and specifications. It will focus on maintaining the document's structure (such as sections for Overview, Requirements, Design, Testing, etc.) and using precise technical language.
Meeting Transcripts
Transcripts capture spoken conversations, which can be informal and unstructured. A protocol tailored to transcripts will likely include steps for identifying different speakers, summarizing key discussion points, and extracting action items or decisions. The style might shift to be more narrative, highlighting timeline of discussion and outcomes (e.g., "Speaker A raised Issue X, and the team decided Y").
Other Documents
Deep Research is flexible and can adapt to many formats, from academic research papers to news articles or emails. In each case, providing a sample will help the AI adjust its approach — whether that means adopting a more scholarly tone for research articles (with proper citations and methodology) or a more concise summary style for emails and memos.
Each document type has unique characteristics. The Deep Research protocol, once adapted, will explicitly account for those nuances in its guidelines. This ensures that the analysis or output you get is context-aware and relevant to the material you're dealing with.

How to use the output: Simply save the new protocol text and upload it with the target document for graph extraction in the deep research UI.

You can use it for analyzing similar documents in the future.

You can copy and paste it into your knowledge base, share it with colleagues, or refer to it whenever you use the Deep Research tool on that type of document.
In short, the tool does the heavy lifting — producing a polished protocol variation — so you can focus on understanding and using the results.
Final Tips for Tweaking the Protocol
Be Specific in Your Request
Clearly state the document type or any particular focus. For example, "adapt the protocol for technical research papers in biology" or "adapt the protocol for financial quarterly reports." The more context you give, the better the AI can fine-tune the guidelines.
Review the Output
Once you receive the adapted protocol, skim through it. It should have logical headings, concise instructions, and formatting appropriate to your document type. If something seems off or you expected a different emphasis, you can clarify or provide additional examples and ask for a refinement.
Reuse and Share
The adapted protocol is meant to be reused. Feel free to share it with your team or include it in your documentation. Next time you analyze a similar document with Deep Research, you can expect the process to follow this custom protocol.
Extracting Your Knowledge Graph
With your tailored protocol ready, it's time to put Deep Research into action. Follow these simple steps to extract a comprehensive knowledge graph from your documents:
Access Deep Research
Navigate to the chat interface and select the Deep Research tool. This is your gateway to advanced document analysis.
Upload Documents & Protocol
Upload your original document (and its Markdown version if available) alongside your custom-tweaked protocol file. Ensure all files are ready for analysis.
Initiate Extraction
Prompt Deep Research with: "Follow the protocol in [your_protocol_filename.md] to extract a knowledge graph from the documents." Be specific with your filename.
Deep Research may ask for clarification to refine the extraction process, ensuring optimal results. This interaction helps fine-tune the output to your exact needs.

It usually takes about 15 minutes for Deep Research to complete. It should output the graph in json-ld format along with a quality assurance report. Review the report and if there are any deficiencies in the result prompt Deep Research to correct the using the protocol as a guide.

We can always come back to this chat and improve the results after some further testing.
Saving Your Knowledge Graph
Standard Format
Your knowledge graph will be exported as a .jsonld file, ensuring a standardized, machine-readable format for broad compatibility with various tools and platforms.
Intuitive Naming
Files are named <document_name>_kg.jsonld, directly linking the graph to its source (e.g., QuarterlyReport_2025_kg.jsonld) for easy identification.
Organized Storage
Always store the .jsonld file alongside its original document and any Markdown versions in the same folder for quick access and coherence within your project files.
Graph Preservation
JSON-LD preserves the intricate graph structure and data provenance, making it an ideal format for direct ingestion by any LLM or advanced analytical tool.
Using Your Knowledge Graph with AI
When you want an AI assistant (like ChatGPT, Gemini, Claude, or Deep Research) to have a deep understanding of your documents, upload three essential files together:
  • Original Document: Your source file (e.g., PDF, DOCX).
  • Markdown Conversion: If you've created one from the original.
  • Knowledge Graph JSON-LD: The structured data file (<document_name>_kg.jsonld) generated by Deep Research.

Most chat interfaces allow you to drag-and-drop multiple files in one message. Upload all three at once for optimal results and richer AI interaction.
Once these files are uploaded, you can ask precise questions:
  • "Summarize the main findings of this document."
  • "What methods are linked to Task A in this report?"
  • "Compare Section 2 of Doc 1 with Section 3 of Doc 2."
The AI assistant will then combine the raw text with the structured knowledge graph to provide sharper, more accurate, and contextually rich answers.

If the AI assistant makes mistakes or gives answers that are incomplete, go back to the Deep Research chat where you created the knowledge graph. Paste in the prompts and poor quality answers from the AI assistant and explain what is missing or incorrect. End with an instruction to improve the graph so that it includes the needed information. Deep Research will use the protocol and your feedback to make the improvements.
Building a Custom LLM Knowledge Base
Empower your AI assistant with deep, contextual understanding by building a custom knowledge base from your documents and their Deep Research graphs.
Single Document Sets
Upload one document (e.g., PDF, DOCX) along with its generated knowledge graph (.jsonld file). This provides focused, in-depth insights on that specific content.
Multiple Document Sets
Combine several documents and their respective graphs to create a comprehensive mini-knowledge base. The LLM can then cross-reference concepts, compare data, and provide richer context across all files.
This allows your LLM to become an instant expert on your specific document collection, enabling advanced queries like comparing quarterly reports or building AI-driven FAQs.
Building a Large Knowledge Base
At some point you will have too many documents and knowledge graphs to fit into the context of Deep Research or other AI assistants.

You can access arbitrarily large graphs using tools such as Graphiti and Neo4j databases. These can be run locally or in the cloud.

We are currently testing a system built with these tools that can support 5-10 users and 10,000 plus documents.

Subscribe for updates to be invited to the next phase of testing.
Start Using Deep Research Today
Looking to supercharge your research process? Discover how our protocol seamlessly integrates with cutting-edge Deep Research tools from leading AI platforms. Our protocol contains detailed instructions for the AI that will allow it to customize the protocol document to match your preferred platform and document formats.
Dive into the resources above to unlock the full potential of Deep Research for your specific needs.