HOW-TO · DEV

How to allocate token budgets across system prompts, context window, and responses

intermediate15 minBy Eruo Fredoline
Target environment
Ubuntu 24.04 · Ollama 0.4.x
PREREQUISITES

Understanding of model context limits and basic familiarity with the target LLM API configuration

What this does

This guide describes a strategy for partitioning a model's total token context window into three functional buckets: system prompt (instructions), conversation history (context), and response generation (output). Allocating these budgets correctly prevents truncation, reduces unnecessary API costs, and keeps model behavior aligned with the intended task.

Steps

  1. Determine the total context window size for the target model. For example, a 4096-token context model has 4096 slots for all input and output combined.
  2. Define a fixed system prompt budget. Reserve 200–500 tokens for the system prompt depending on instruction complexity. Store this value in a named constant.
  3. Define a fixed response budget. Reserve 256–1024 tokens for the expected response length, depending on task type. This prevents the model from consuming all remaining tokens with a single reply.
  4. Calculate the available conversation history budget: context_window - system_budget - response_budget. For a 4096 window with 400-system and 512-response budgets, 3184 tokens remain for history.
  5. Encode conversation history into the prompt, truncating from the oldest messages first when the budget is exceeded.
  6. Instrument the pipeline to log actual token counts for each category during development to validate allocations.
  7. Adjust budgets per task type: longer analytical tasks warrant higher response budgets; short Q&A tasks can allocate more to history.
  8. Set soft limits in the API call (e.g., max_tokens parameter) to enforce the response budget hard stop.

Verification

python3 -c "
context = 4096
system_budget = 400
response_budget = 512
available = context - system_budget - response_budget
print(f'Available for history: {available} tokens')
"
# Expected output: Available for history: 3184 tokens

Common failures

  1. History overflow: Conversation history exceeds the budget, causing truncation of recent messages and loss of critical context. Solution: implement a ring buffer that drops the oldest messages first, preserving the most recent exchanges.
  2. Response truncation: Setting max_tokens too low causes the response to be cut mid-sentence. Solution: estimate required response length from task type and set max_tokens to at least estimated_output * 1.5.
  3. System prompt bloat: Adding verbose instructions to the system prompt consumes tokens that should belong to conversation history. Solution: keep system prompts concise; extract detail into separate documents referenced by the system prompt.
  4. No budget monitoring: Pipeline runs without token counting, leading to unpredictable truncation. Solution: add per-request token logging in development mode.

Related guides

← All how-to guidesCourses →