API Best Practices

A rough rule of thumb: 1 token equals ~4 characters of English text, or ~1-2 Chinese characters. For code, 1 token is typically 2-4 characters. Most chat messages of a few sentences consume 100-500 tokens depending on length and language.

If you're building a cost-sensitive application, count tokens before sending requests using tiktoken or a similar library. OpenAI's tokenizer also works well for DeepSeek since both use similar tokenization schemes.

System prompts are powerful but they consume input tokens on every request. Keep them concise. If you have a complex, reusable instruction set, consider:

1. Using a short system prompt with specific task instructions in the user message
2. Summarizing long conversations before sending to the API
3. Using cache-hit tokens for repeated patterns in system prompts

Example efficient system prompt:

You are a concise technical assistant. Respond directly. No preamble.

DeepSeek imposes rate limits per token. When you hit HTTP 429, implement exponential backoff:

import time, random
for attempt in range(5):
response = call_api(...)
if response.status_code != 429:
break
wait = (2 ** attempt) + random.uniform(0, 1)
time.sleep(wait)

Never spin-wait in a loop. Always use exponential backoff with jitter.

deepseek-chat (v4-flash): General-purpose chat. Fast, cost-effective. Use for: code generation, content writing, Q&A, summarization, translation, most tasks.

deepseek-reasoner (v4-flash reasoning): Extended chain-of-thought reasoning. Slower but more thorough. Use for: complex multi-step problems, mathematical proofs, strategic analysis, debugging tricky issues.

For most production applications, deepseek-chat is the better default choice.

DeepSeek supports up to 1M token context windows, but longer contexts cost more and can reduce response quality. For ongoing conversations:

1. Periodically summarize and truncate conversation history
2. Use sliding window approaches (keep last N messages)
3. For retrieval-augmented tasks, fetch relevant context on-demand rather than preloading everything

This keeps costs predictable and improves model focus.

Common error codes and how to handle them:

• HTTP 400 Bad Request: Malformed JSON or invalid parameters. Check your request body.
• HTTP 401 Unauthorized: Invalid or expired token. Check your API key.
• HTTP 429 Too Many Requests: Rate limit hit. Back off and retry.
• HTTP 500/502/503 Server Error: DeepSeek-side issue. Retry after a delay.

Always log errors with full context (request parameters, error body) for debugging.