Published

GenAI Service Audit

Goulven
Goulven

GenAI Service Audit

Architecture Overview

The GenAI service in Open4Goods is a distributed system involving the Frontend, Front-API, and the Back-Office API (Core).

Flow Diagram

Quota Management

IP-based quotas are enforced at the Front-API level to protect the back-office from abuse.

  • Service: QuotaService
  • Configuration: ReviewGenerationProperties (default: 3 requests per IP per 24h).
  • Enforcement: ProductMappingService checks QuotaService before calling the ReviewGenerationClient.
  • Frontend: AccountPrivacyCard.vue uses useIpQuota.ts to display remaining quota to the user.

Prompt Engineering

The system uses a "POJO-oriented" prompt engineering approach, combining external templates with code-defined instructions.

  1. Template (review-generation.yml): Defines the base System and User prompts. It uses Thymeleaf for variable substitution (e.g., ${PRODUCT_NAME}, ${sources}).
  2. Code Instructions (@AiGeneratedField): specific instructions for each field of the AiReview object are defined directly in the Java class using annotations. 
    @AiGeneratedField(instruction = "Description du produit, 150 mots maximum")
private String description;
  3. Concatenation (PromptService):
    • Calls AiFieldScanner.getGenAiInstruction(type) to extract all annotated instructions.
    • Appends these instructions to the System Prompt:

      "En complément du schéma JSON, voici les instructions concernant chaque champs que tu dois fournir…"

Benefits & Risks

  • Benefits: Keeps instructions close to the data model (Documentation as Code). Ensures the AI knows exactly what to generate for each JSON field.
  • Risks:
    • Localization: Hardcoded French text in PromptService limits specific support to other languages.
    • Determinism: Field order depends on reflection (getDeclaredFields), which isn't guaranteed deterministic without explicit sorting.

ReviewGenerationService Analysis

The ReviewGenerationService orchestrates the complex flow of gathering context before calling the AI.

Workflow

  1. Check: shouldGenerateReview (cache validity).
  2. Preprocessing:
    • Search: Google Search for Brand + Model.
    • Fetch: Concurrent scraping of top results.
    • Filter: Discard PDF, low token count, or preferred domain logic.
    • Count: Accumulate tokens up to maxTotalTokens.
  3. Prompt Prep:
    • Injects sources, tokens, PRODUCT_DETAILS.
    • Manually constructs the ATTRIBUTES list string.
  4. Generation: Calls PromptService with "review-generation" key.
  5. Post-processing:
    • updateAiReviewReferences: Replaces [1] with HTML links.
    • populateAttributes: Extracts dynamic attributes.