Every number, name and error code worth memorising

Every constant scattered through the previous cards, collected. Nothing here needs understanding — it needs recall, so read it last and read it often.

The numbers

Vector store and files

• 10,000 files per vector store
• 512 MB per file, roughly 5 million tokens
• Chunk 800 tokens, overlap 400
• 20 chunks maximum in context
• text-embedding-3-large, 256 dimensions

Requests and media

• 10 images per request
• 25 MB maximum audio request
• Video seconds accepts 4, 8 or 12
• One token is roughly 4 to 8 characters
• Cost is quoted per million tokens

Scores and thresholds

• Retrieval evaluators score 1 to 5, passing at 3
• Confidence runs 0 to 1; 0.9 or above is safe for straight-through processing
• Semantic reranker scores 0 to 4, over the top 50
• Vector similarity runs 0.333 to 1.0
• Fusion constant k is about 60
• The exam pass mark is 700 out of 1000

Time and size limits

• Runs expire after 10 minutes
• Batch turnaround is around 24 hours
• A virtual network subnet must be /27 or larger
• The content filter default threshold is medium

The error codes

Code	What it means	What to do
429	Rate-limited	Exponential backoff plus jitter; honour Retry-After
403	Quota window exceeded	Request more quota
400 with content_filter	The prompt was blocked	Rewrite the prompt
200 with finish_reason: content_filter	The completion was filtered	Handle partial output
410 Gone	The model is retired	Migrate to the replacement
403 on botServices/write	Missing Azure Bot Service Contributor	Assign the role

Names that must be spelled correctly

azure-ai-projects · AIProjectClient · responses.create · DefaultAzureCredential · TextAnalyticsClient · ContentUnderstandingClient · SpeechRecognizer · SpeechSynthesizer · SpeechTranslationConfig · azure-ai-voicelive · Foundry User, the inference role.

The trap

429 and 403 look interchangeable and are not.

429 is rate-limiting — you are going too fast at this moment. 403 is the quota window — you have exhausted your allowance.

The tell is in how the scenario describes the overuse. Sustained overuse across a billing period wants 403. A burst wants 429. Once you have that, the remedy follows: a burst is fixed by backing off, an exhausted allowance is not fixed by waiting a few seconds.