Personalised Photo Book Title Generation Using Retrieval-Augmented LLMs

popsa · Tech · 2026 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Tech

Popsa wanted to generate better photo-book titles and subtitles for users who often chose generic titles, while meeting strict output format and brand requirements across languages.

Constraints:

• titles and subtitles must meet character and category rules
• outputs must be valid JSON
• suggestions need to be brand-aligned and multilingual
• quality, cost, and latency all matter for a customer-facing feature

Architecture

Popsa's Title Suggestion service starts inside the mobile photo-book workflow: the app reads timestamps/geocoordinates and runs on-device CNNs to identify image features, then the service decrypts/processes the design, reverse-geocodes locations, classifies the subject, and turns the result into a compact book description. A retrieval-based few-shot component fetches similar Photo Book examples and accepted titles from an example database, seeds an Amazon Bedrock conversation with those prior user/assistant examples, and asks Claude or Nova models for valid JSON containing title, subtitle, and category. Later latency work moved from Bedrock InvokeModel to ConverseStream, with FastAPI parsing streaming title-subtitle-category triplets and returning the first valid suggestion immediately while more suggestions continue in the background.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  mobile-photo-book-app["mobile-photo-book-app"] --> on-device-vision["on-device-vision"]
  mobile-photo-book-app --> title-suggestion-service["title-suggestion-service"]
  on-device-vision --> title-suggestion-service
  title-suggestion-service --> reverse-geocoder["reverse-geocoder"]
  title-suggestion-service --> subject-classifier["subject-classifier"]
  title-suggestion-service --> retrieval-few-shot-prompter["retrieval-few-shot-prompter"]
  example-photo-book-database["example-photo-book-database"] --> retrieval-few-shot-prompter
  retrieval-few-shot-prompter --> amazon-bedrock["amazon-bedrock"]
  amazon-bedrock --> streaming-fastapi-parser["streaming-fastapi-parser"]
  amazon-bedrock --> evaluation-loop["evaluation-loop"]

Components

Data flow

A customer selects photos for a book. The mobile app extracts timestamps/geocoordinates and on-device image features, then the Title Suggestion Service decrypts and processes the design, reverse-geocodes coordinates, classifies the design subject, and builds a natural-language description. That description enters retrieval-based few-shot prompting: similar historical photo books and accepted titles are retrieved, formatted as prior conversation examples, and sent through Amazon Bedrock to Claude or Nova. The model returns JSON title/subtitle/category suggestions. In the streaming version, FastAPI validates streamed triplets and returns the first polished suggestion as soon as it is valid.

Why this architecture

The architecture keeps deterministic product constraints close to the service and uses generative AI only after the design has been summarized into grounded book context. Retrieval-based few-shot prompting won because similar accepted examples let the LLM emulate brand/style while following JSON and category rules. Bedrock's unified API made model-family comparisons cheap, and ConverseStream solved perceived latency by returning the first validated suggestion before the full batch finished.

Models used: Claude 3 Haiku · Amazon Nova Lite · Amazon Nova Pro · Amazon Nova Micro · on-device convolutional neural networks

Scale

• Popsa operates across more than 50 countries and 12 languages.
  “Available across more than 50 countries and 12 languages, we use design automation and AI to transform everyday photos into personal, shareable experiences”article, stated
• The evaluation dataset contained over 100 example Photo Books.
  “we built a dataset of over 100 example Photo Books and defined our metrics in an evaluation pipeline”article, stated
• A/B testing ran with hundreds of thousands of users before full rollout.
  “We also conducted multivariate testing with hundreds of thousands of users.”article, implied
• The feature has generated over 5.5 million personalized titles in 2025.
  “In 2025, our Title Suggestion feature has generated over 5.5 million titles”article, stated
• ConverseStream reduced average time to first suggestion from 1.41 seconds to 0.92 seconds, a 35% improvement.
  “we reduced the average time to first suggestion from 1.41 seconds to 0.92 seconds, delivering title suggestions 35% faster to users.”article, stated

Alternatives considered (roads not taken)

The rule/template Title Suggestion Graph

Rejected because: Retrieval-based few-shot prompting produced more creative, higher-rated suggestions than the earlier graph algorithm.

Amazon Nova Micro for production title generation

Rejected because: It beat the legacy graph but lagged other LLMs in user satisfaction, so Popsa focused on the remaining models.

Claude 3 Haiku as the ongoing default

Rejected because: Nova Lite offered nearly identical quality at lower cost and faster response times, while Nova Pro offered the highest satisfaction at higher cost.

Waiting for a full batch through the Bedrock InvokeModel API

Rejected because: Streaming via ConverseStream lets the service return the first valid suggestion immediately, improving perceived responsiveness.

Patterns

rag-retrieval (retrieval)

For a new photo book, the system retrieves similar examples and their accepted title suggestions as few-shot context.

structured-output-contract (output)

The model is constrained to return valid JSON with title, subtitle, and category keys.

model-routing (model-operations)

Popsa used Bedrock's unified API to compare Claude and Nova model families by changing model IDs.

Tradeoffs

Move from rules/templates to retrieval-augmented generative suggestions.

Chose retrieval-based few-shot prompting over the Title Suggestion Graph rule/template algorithm.

Rationale: The source says the earlier graph used rules and templates, while top results came from retrieval-based few-shot prompting.

Prioritize Nova Lite for production economics after model comparison.

Chose Amazon Nova Lite for near-Haiku quality at lower cost and faster response over staying only with Claude 3 Haiku or using Nova Pro everywhere.

Rationale: Nova Pro had the highest user satisfaction, but Nova Lite gave near-identical quality to Claude Haiku with better cost and latency.

Stream and validate suggestions incrementally.

Chose ConverseStream with real-time FastAPI parsing over waiting for the entire batch from InvokeModel.

Rationale: Returning the first validated triplet made the feature feel responsive while additional suggestions continued in the background.

Evaluation

Popsa used programmatic guardrail checks, LLM-based quality assessment, internal testing, A/B tests, and user feedback.

Metrics: valid JSON rate · category validity · character-limit compliance · positive user feedback · cost · latency

Outcomes

• Generative AI titles increased positive user feedback by 13 percentage points over the graph algorithm.

The source reports over 5.5 million personalized titles generated in 2025.