Deep Research Experiment

NotebookLM allows you to ask questions of specific sources and receive answers based strictly on those sources. But what do you do if the sources in your notebook don’t have an answer? That’s where Deep Research comes in.

Over the next three days I’m going to publish articles about why Upāli, who was famous for observing and keeping the monastic precepts, is missing from the Lotus Sutra and about the Diamond Chalice Precept, the single precept that replaced the monastic code. These were generated from the “Deep Research” option in Gemini and in NotebookLM.

The articles The Hermeneutics of Absence: Upāli and the De-Emphasis of Monastic Legalism in the Lotus Sūtra and Diamond Chalice Precept: Scriptural Genesis, Ontological Evolution, and Hermeneutical Implementation in East Asian Buddhism were generated using Google’s Gemini Deep Research feature and the article The Indestructible Vow: The Diamond Chalice Precept (Kongō-hōki-kai) was generated as a report in NotebookLM.

Both Gemini and NotebookLM do a great job finding sources to answer a specific question, digesting the information and returning a reasoned report. The big difference between Deep Research in Gemini vs. NotebookLM is the inclusion of citations linking the text to its sources. (NotebookLM does not include footnotes in its reports since you already know the sources. ) The problem for me was how to get those citations into the articles I publish on 500yojanas.org.

Complicating everything is that multiple citations can be used for a particular text. For example:

Activated through faith and the contemplation of the ultimate Dharma, this single precept is understood to encompass and fulfill all other moral rules within itself.25

When I had Gemini convert the Upali report to HTML, it selected just one citation to keep. I don’t know whether that was random or reasoned. For the Diamond Chalice report, Gemini kept all of the citations and provided links back from the footnotes to the article. Here’s an example of what that looks like:

  1. “The Diamond Chalice Precept in East Asian Buddhism,” Tricycle Magazine. Back to text:
    5.1 | 5.2 | 5.3 | 5.4 | 5.5 | 5.6 | 5.7 | 5.8 | 5.9 | 5.10 | 5.11 | 5.12 | 5.13 | 5.14

The fact that one article has a single citation and the other has return links is a good example of the randomness of Gemini’s output if you don’t specifically tell it what you want. To get a consistent output I need to give Gemini an example of what I want and explicit instructions. Here’s what works for me:

Convert this document into an HTML document that can be copied and pasted into a classic editor in WordPress. Use the example-code.txt as a source for how to create the links from the body of the text to the sources and back again.

Format Conversion: Translate all Markdown text, tables, blockquotes, and lists into equivalent, clean, and compliant HTML blocks optimized for pasting directly into the WordPress Classic Editor (HTML Tab).

Superscript Citations: Convert all numeric citations (e.g., [1, 2]) to superscript tags containing distinct anchored links (3).

Bidirectional Hypertext Links: Create a Footnotes & Sources list at the bottom matching each numbered citation. Each multi-use footnote features individual, indexed backward links to return the user exactly to where that citation appears in the document text.

Text Preservation: Keep every word, heading, and table entry strictly identical to your original report without changes or paraphrasing.

Going back to the original problem – your NotebookLM notebook doesn’t have sources for your question – the answer is to upload the new report as a source. Alternatively, you can upload some or all of the new sources identified, but that has to be done manually.