LLM

LLM Implementation Guidelines

Directory Structure

LLM-related code is organized in specific directories:

• apps/web/utils/ai/ - Main LLM implementations
• apps/web/utils/llms/ - Core LLM utilities and configurations
• apps/web/__tests__/ - LLM-specific tests

Key Files

• utils/llms/index.ts - Core LLM functionality
• utils/llms/model.ts - Model definitions and configurations
• utils/llms/use-cases.ts - Product use-case to model-role routing
• utils/usage.ts - Usage tracking and monitoring

Model Routing

For product features with a static model choice, use getModelForUseCase(emailAccount.user, LlmUseCase.FeatureName) from utils/llms/use-cases.ts. Keep direct getModel(user, modelType) calls for generic helpers where the model role is intentionally passed from upstream. When adding or changing a use case, update utils/llms/use-cases.test.ts.

Implementation Pattern

Follow this standard structure for LLM-related functions:

import { z } from "zod";
import { createScopedLogger } from "@/utils/logger";
import { chatCompletionObject } from "@/utils/llms";
import type { EmailAccountWithAI } from "@/utils/llms/types";
import { createGenerateObject } from "@/utils/llms";
import { getModelForUseCase, LlmUseCase } from "@/utils/llms/use-cases";

export async function featureFunction(options: {
  inputData: InputType;
  emailAccount: EmailAccountWithAI;
}) {
  const { inputData, user } = options;

  if (!inputData || [other validation conditions]) {
    logger.warn("Invalid input for feature function");
    return null;
  }

  const system = `[Detailed system prompt that defines the LLM's role and task]`;

  const prompt = `[User prompt with context and specific instructions]

<data>
...
</data>

${emailAccount.about ? `<user_info>${emailAccount.about}</user_info>` : ""}`;

  const modelOptions = getModelForUseCase(
    emailAccount.user,
    LlmUseCase.FeatureName,
  );

  const generateObject = createGenerateObject({
    userEmail: emailAccount.email,
    label: "Feature Name",
    modelOptions,
  });


  const result = await generateObject({
    ...modelOptions,
    system,
    prompt,
    schema: z.object({
      field1: z.string(),
      field2: z.number(),
      nested: z.object({
        subfield: z.string(),
      }),
      array_field: z.array(z.string()),
    }),
  });

  return result.object;
}

Best Practices

1. System and User Prompts:

   • Keep system prompts and user prompts separate
   • System prompt should define the LLM's role and task specifications
   • User prompt should contain the actual data and context

2. Schema Validation:

   • Always define a Zod schema for response validation
   • Make schemas as specific as possible to guide the LLM output

3. Logging:

   • Use descriptive scoped loggers for each feature
   • Log inputs and outputs with appropriate log levels
   • Include relevant context in log messages

4. Error Handling:

   • Implement early returns for invalid inputs
   • Use proper error types and logging
   • Implement fallbacks for AI failures
   • Add retry logic for transient failures using withRetry

5. Input Formatting:

   • Use XML-like tags to structure data in prompts
   • Remove excessive whitespace and truncate long inputs
   • Format data consistently across similar functions

6. Type Safety:

   • Use TypeScript types for all parameters and return values
   • Define clear interfaces for complex input/output structures

7. Code Organization:

   • Keep related AI functions in the same file or directory
   • Extract common patterns into utility functions
   • Document complex AI logic with clear comments

8. AI-First Behavior:

   • Prefer generic prompt instructions, structured outputs, and model choice over brittle lexical heuristics that imitate model reasoning
   • Only add deterministic filters when the product truly needs a hard rule outside the model
   • Do not add prompt examples that closely mirror eval fixtures just to make a test pass

9. Draft Attribution Versioning:

   • When changing draft-generation prompt inputs, retrieval context, model routing behavior, or post-processing, bump apps/web/utils/ai/reply/draft-attribution.ts DRAFT_PIPELINE_VERSION
   • Do not bump it for behavior-preserving refactors that keep the same prompt, context, model role, and output processing
   • Treat that version as analytics attribution for reply-draft quality comparisons

Testing

See llm-test.mdc