cohere

Generating Marketing Copy with Cohere and Picking the Best Candidate

A marketing team needs multiple variations of product copy, not just one shot. The system generates three candidates via Cohere's Generate API, then selects the one with the highest likelihood score. This separates prompt construction, multi-generation, and quality-based selection into distinct pipeline stages.

Workflow

productDescription, targetAudience
       │
       ▼
┌──────────────────────┐
│ cohere_build_prompt  │  Assemble generate request body
└──────────┬───────────┘
           │  requestBody
           ▼
┌──────────────────────┐
│ cohere_generate      │  Call Cohere API (3 generations)
└──────────┬───────────┘
           │  apiResponse.generations
           ▼
┌──────────────────────┐
│ cohere_select_best   │  Pick highest-likelihood generation
└──────────────────────┘
           │
           ▼
    bestGeneration, allTexts

Workers

CohereBuildPromptWorker (cohere_build_prompt) -- Builds a LinkedHashMap request body for the Cohere Generate API. Constructs the prompt as "Write a compelling marketing copy for the following product. Target audience: " + targetAudience + ". Product: " + productDescription. Workflow defaults: model "command", maxTokens: 300, temperature: 0.9, k: 0, p: 0.75, numGenerations: 3. Also sets return_likelihoods: "GENERATION" and truncate: "END".

CohereGenerateWorker (cohere_generate) -- Checks COHERE_API_KEY env var at construction. In live mode, POSTs to https://api.cohere.ai/v1/generate with a 10-second connect timeout and 60-second request timeout. In fallback mode, returns 3 hardcoded generations with likelihoods of -1.82, -1.65, and -1.91, plus billing metadata of 42 input / 185 output tokens. The fallback generations are realistic SmartBoard Pro marketing copy variants.

CohereSelectBestWorker (cohere_select_best) -- Iterates through the generations list, comparing each entry's likelihood (a Number cast to double) against Double.NEGATIVE_INFINITY to find the maximum. The least-negative likelihood wins -- in the fallback data, that's -1.65 (the second generation). Also collects all texts into an allTexts list via .stream().map(...).toList().

Tests

15 tests across 3 test files cover prompt construction, API call modes, and best-candidate selection logic.

Further Reading

• RUNNING.md -- how to build and run this example