Accessing the API

Full Request Lifecycle (5 Steps)

• Step 1: User enters text in web app and clicks send
• Step 2: Client sends message to your backend server (never call API directly from client-side)
• Step 3: Server makes request to Anthropic API using SDK or HTTP
• Step 4: API generates text and returns response
• Step 5: Server sends generated text back to client for display

API Request Requirements

• Must include secret API key (keep on server only, never expose in client code)
• Must specify model name
• Must provide list of messages (including user input)
• Must set max tokens value (limits response length)

Text Generation Process (4 Stages)

Stage 1 - Tokenization:

• User input broken into tokens (words, word parts, spaces, numbers)

Stage 2 - Embedding:

• Each token converted to embedding (list of numbers representing word meaning)
• Single word can have multiple meanings depending on context

Stage 3 - Contextualization:

• Each embedding adjusted based on surrounding embeddings
• Refines meaning to match context and neighboring words

Stage 4 - Generation:

• Output layer produces probabilities for next possible word
• Model uses mix of probability and randomness (not just highest probability)
• Selected word added to sequence and process repeats

Generation Stopping Conditions

• Token count exceeds max tokens parameter
• Model generates special end-of-sequence token (signals natural completion)

API Response Contains

• Generated message text
• Usage data (input tokens + generated tokens count)
• Stop reason (why generation ended)

Takeaways

• Always route API requests through your backend server using the API key stored there—never call the API directly from client-side code
• API requests require the model name, message list, max tokens limit, and secret API key
• Text generation works in four stages: tokenization → embedding → contextualization → generation, with the model selecting words based on probability plus randomness rather than just picking the highest probability
• Generation stops when max tokens is reached or the model outputs an end-of-sequence token
• API responses include the generated text, token usage counts, and a stop reason explaining why generation ended