Reverse Engineering the Response to the prompt "best smartphones 2025" in Google Gemini

      By Chris Byrne (7.8.25)

     This is my (latest) attempt to reverse engineer a response to a comparative entity prompt ("best smartphones 2025") in Google Gemini using the tool itself. I chose this prompt to trigger "Retrieval Augmented Generation" (RAG): where instead of possibly relying on the model’s pre-trained knowledge alone, the model would be most likely to retrieve 'current' relevant information from connected data sources.

    Here was the response:

“Best Smartphones of 2025

    The 2025 smartphone market is defined by a fierce competition in AI integration, camera technology, and the continued maturation of foldable phones. The "best" phone is a subjective choice, but here are the top-rated devices across key categories:

•  Best Overall Flagships: The Samsung Galaxy S25 Ultra and Apple iPhone 16 Pro Max are widely considered the two dominant forces. The S25 Ultra is praised for its all-around excellence, including the S Pen stylus and powerful AI features, while the iPhone 16 Pro Max is lauded for its seamless user experience and powerful A18 Pro chip, which powers Apple Intelligence features .
  

• Best for Photography: For photography enthusiasts, the Google Pixel 9 Pro is considered the ultimate camera phone for its computational photography that delivers true-to-life images with excellent dynamic range and color reproduction. The Samsung Galaxy S25 Ultra is a close contender, especially for its impressive 5x telephoto zoom and new 50MP ultrawide camera . The iPhone 16 Pro Max is also a top pick for its natural color reproduction and advanced video capabilities.
  

• Best for Gaming: For dedicated mobile gamers, the Asus ROG Phone 9 Pro is the top choice, built with a 185Hz AMOLED screen, a powerful Snapdragon 8 Elite chipset, and physical Air Trigger controls. Mainstream options like the Samsung Galaxy S25 Ultra and iPhone 16 Pro Max are also excellent all-rounders, handling demanding games with ease thanks to their top-tier processors .
  

• Best Budget and Mid-Range Phones: For those seeking value, the Google Pixel 9a stands out as an excellent midrange option with a capable camera and smooth performance [8]. Other strong budget-friendly choices include the Samsung Galaxy A16 5G and Motorola Moto G Power 5G (2024), which offer solid performance and impressive battery life for their price points ."
  
  

My analysis (assisted by Gemini)

    Here is how Google Gemini stated the response was formed:

1. Data Pipeline and Information Gathering:

Gemini stated that the response was built using a retrieval-based and extractive (R.E.) data pipeline. The process began with targeted web searches to gather current and specific information on the 2025 smartphone market.

A. Exact Web Searches Used (and Search Engine):

Gemini stated that all information was gathered using the Google search engine using the following queries (thus no pre-existing knowledge from it's corpus was used):

1. "best smartphones of 2025 review"

2. "Samsung Galaxy S25 Ultra review PCMag"

3. "Google Pixel 9 Pro camera review"

4. "best gaming phone 2025 specs"

5. "best budget phones 2025 TechRadar"
   
   

B. Synthesis and R.E. Embeddings (Retrieval-Augmented Embeddings):

   Gemini stated that the search results from these queries were processed, and key facts were extracted and "embedded" into a knowledge base. These R.E. embeddings allowed for the synthesis of a comprehensive response by pulling specific data points from multiple sources: (note I have broken some URLs below)

• [1, 2] Best Flagships: Gemini stated that the consensus among sources like PCMag (https://www.pcmag .com/picks/the-best-phones) and ZDNet (https://www.zdnet  .com/article/best-phone/) was that the Samsung Galaxy S25 Ultra and iPhone 16 Pro Max are the top-tier phones. The S25 Ultra's S Pen and AI features were highlighted by PCMag, while the iPhone's A18 Pro chip and Apple Intelligence were a focus of ZDNet's review.
  

• [3, 4, 5] Photography: Gemini stated that it used a combination of reviews from DxOMark (https://www.dxomark .com/google-pixel-9-pro-xl-camera-test/), GSMArena (https://www.gsmarena .com/google_pixel_9_pro-review-2745p5.php), and ZDNet (https://www.zdnet .com/article/best-phone/) to determine the best camera phones. The Pixel 9 Pro's computational photography was praised by DxOMark, while the S25 Ultra's hardware and zoom capabilities were highlighted by TechRadar and ZDNet. The iPhone's video and natural color were also mentioned as a strength in these reviews.
  

• [6, 7] Gaming: Gemini stated that it used reviews from GamesRadar (https://www.gamesradar .com/best-phones-for-gaming/) and PCMag (https://www.pcmag   .com/picks/the-best-gaming-phones) consistently named the Asus ROG Phone 9 Pro as the top dedicated gaming phone due to its specialized hardware, with ZDNet also acknowledging the S25 Ultra's Snapdragon 8 Elite and the iPhone 16 Pro Max's A18 Pro as excellent for mainstream gaming.
  

• [8, 9] Budget Phones: Gemini stated that it used reviews from publications like PCMag (https://www.pcmag .com/picks/the-best-phones) and Expert Reviews (https://www.expertreviews .co.uk/technology/phones/best-budget-smartphones) to identify the best budget phones. The Google Pixel 9a was cited for its midrange value, while the Samsung Galaxy A16 5G and Motorola Moto G Power 5G were identified as strong, entry-level choices.”
  
  

Embedding extraction for R.E. embedding process

   I also asked Chatgpt to outline embedding extraction for R.E. embedding process for above "best smartphones 2025" type-prompt :Chatgpt stated, “The embedding extraction for the R.E. (Retrieval-Augmented) embedding process, specifically for the "best smartphones 2025" prompt, involves a series of steps to convert raw web data into a searchable, numerical format. Here's a detailed outline of that process:

1. Data Source Identification and Acquisition

   The process begins with the identification of relevant data sources. For the "best smartphones 2025" prompt, this involves fetching content from high-authority tech review websites and publications. The sources would include the web pages from the searches e.g. 

• Source 1: https://www.pcmag .com/picks/the-best-phones
  

• Source 2: https://www.techradar .com/phones/samsung-galaxy-phones/samsung-galaxy-s25-ultra-review
  

• Source 3: https://www.zdnet .com/article/best-phone/
  

• Source 4: https://www.gamesradar .com/best-phones-for-gaming/
  

...and so on.

2. Text Chunking and Preprocessing

The fetched HTML content from these URLs is then processed as follows:

•  Extraction: Irrelevant elements like ads, navigation bars, and footers are stripped away to isolate the main body of the article.
  

• Chunking: The remaining article text is divided into logical chunks. A common strategy for this specific content would be to chunk by paragraph or by a section heading. For example, a review article would be broken into chunks e.g. 

Chunk A: "Best Overall Flagships" section
Chunk B: "Best Camera Phone" section
Chunk C: A paragraph detailing the "Samsung Galaxy S25 Ultra's" AI features.
Chunk D: A paragraph about the "Google Pixel 9 Pro's" computational photography.

3. Vectorization (Embedding)

Each text chunk is independently converted into a numerical vector:

• Embedding Model: A pre-trained language model, like a variant of BERT or a sentence transformer, is used. This model has been trained on a massive amount of text to understand the semantic relationships between words and phrases.
  

• Conversion: The text from Chunk C ("A paragraph detailing the Samsung Galaxy S25 Ultra's AI features...") is fed into the embedding model. The model processes the text and outputs a dense vector, which is a list of several hundred numbers. This vector is a numerical representation of that chunk's specific meaning and content.
  
  

4. Indexing and Storage

The generated embeddings are then stored in a vector database for efficient retrieval:

•  Storage: Each vector is stored alongside its original text chunk and its metadata (e.g., source URL, date of publication, etc.).
  

• Indexing: The database creates an index of these vectors, which is a data structure optimized for fast similarity searches. This allows the system to quickly find the most relevant chunks without having to compare a new query against every single vector.
  
  

5. Retrieval for Response Generation

When a user submits the prompt "best smartphones 2025," the R.E. process leverages this prepared knowledge base:

•  Query Embedding: The user's query itself is converted into an embedding using the same model used for the source data.
  

• Similarity Search: The system performs a similarity search in the vector database to find the vectors that are "closest" to the query vector.
  

• Retrieval: The database returns the "top N" most relevant text chunks  (taking the "N" most likely next tokens i.e. words, subwords, or characters from the model’s probability distribution when generating text - "N" is the count of the most probable tokens the model is allowed to pick from at each generation step ) . For this prompt, it would likely retrieve chunks detailing the Samsung Galaxy S25 Ultra, the iPhone 16 Pro Max, the Google Pixel 9 Pro, and so on, as these chunks are semantically most similar to the query.
  

• Synthesis: These retrieved chunks of text are then used to augment the language model's response generation. The model synthesizes the information from these specific, factual chunks to produce the final, comprehensive ... response.”

I did the same in Chat GPT too.