In `@backend/src/api/routes/ai/index.routes.ts`:
- Around line 477-513: Add a rate limiter middleware to the embeddings route:
create an express-rate-limit instance (e.g., embeddingsRateLimiter using
rateLimit with sensible windowMs/max settings) and insert it into the middleware
chain for the router.post('/embeddings') route (for example:
router.post('/embeddings', embeddingsRateLimiter, verifyUser, ...)). Ensure you
import rateLimit from 'express-rate-limit', configure the limiter, and apply it
before the handler that calls EmbeddingService.getInstance().createEmbeddings to
enforce limits on this cost-heavy endpoint.

In `@backend/src/services/ai/ai-usage.service.ts`:
- Around line 93-121: In trackEmbeddingUsage, stop computing outputTokens when
either totalTokens or inputTokens is undefined and ensure outputTokens is
clamped to >=0 so zeros are preserved; specifically, change the logic in
trackEmbeddingUsage to: only compute outputTokens when both inputTokens and
totalTokens are numbers, set outputTokens = Math.max(0, totalTokens -
inputTokens), and pass null for input_tokens/output_tokens in the INSERT only
when the corresponding value is undefined (not when it's 0), updating the values
array and logger usage accordingly.

In `@backend/tests/manual/test-ai-embeddings.sh`:
- Line 219: The script always prints "print_success" and exits 0 even when
track_test_failure sets TEST_FAILED; update the end of the test (around the
print_success "🎉 AI Embeddings test completed!" call) to check the TEST_FAILED
variable and if set exit with a non-zero status (e.g., call print_failure or
echo an error and exit 1), otherwise print_success and exit 0; reference the
track_test_failure function and TEST_FAILED variable to locate where to add this
conditional exit.

In `@shared-schemas/src/ai-api.schema.ts`:
- Around line 159-189: The embeddingObjectSchema currently forces embedding:
z.array(z.number()) which breaks for encoding_format="base64"; change
embeddingObjectSchema to use a union for embedding:
z.union([z.array(z.number()), z.string()]) so embeddings can be either number[]
or base64 string, and update any code that maps/normalizes responses (the
service mapping that currently does an unsafe "as number[]" cast) to stop
casting—preserve the embedding field as number[] | string and add explicit
handling: if embedding is a string decode base64 to numbers in a dedicated
helper, otherwise pass through the number[]; update signatures/types accordingly
so embedding is typed as number[] | string (symbols: embeddingObjectSchema,
embeddingsResponseSchema and the service mapping that performs the unsafe cast).

In `@backend/src/services/ai/embedding.service.ts`:
- Line 4: The import for AIConfigService in embedding.service.ts is missing the
.js extension which breaks ESM resolution; update the relative import statement
that references AIConfigService (symbol: AIConfigService) to include the .js
extension (e.g., change './ai-config.service' to './ai-config.service.js') so it
matches other imports like './ai-usage.service.js' and '@/utils/logger.js' and
ensures consistent ESM module resolution at runtime.

In `@backend/src/services/ai/embedding.service.ts`:
- Around line 56-68: The embedding service is incorrectly calling
this.aiUsageService.trackChatUsage(...) and using chat-oriented token math;
replace that with a dedicated embedding tracking call: add a new method on
AiUsageService (e.g., trackEmbeddingUsage(aiConfigId, inputTokens, model,
metadata?)) and call it from embedding.service.ts instead of trackChatUsage,
passing tokenUsage.promptTokens as inputTokens and a fixed 0 for outputTokens
(or omit output tokens), and ensure the new method persists a
usage_type='embedding' (or adds a usage_type column in the ai_usage schema) so
embedding usage can be distinguished in analytics/billing; update any
callers/tests accordingly to use trackEmbeddingUsage and options.model for the
model id.