context-optimization

Context Optimization

The Law

PROMPT COST IS NOT OPTIMISED BY GUESSING.
"I already have a short prompt" is a guess about token count.
"Reducing context will hurt quality" is a guess about the quality/cost curve.
Measure first. Apply the hierarchy. Measure again. THEN claim improvement.

When to Use

Trigger when:

• Prompt + context exceeds 60% of the model's context window
• p95 latency is above target
• Cost per 1k calls exceeds budget
• RAG pipeline is stuffing too many chunks into each call
• Output quality is degrading on long inputs (long context dilution)

When NOT to Use

• The feature has < 100 calls/month and total monthly cost is < $20 — optimisation ROI is negative at this volume
• Quality is currently below threshold — fix quality first; optimising a broken pipeline only makes it cheaper to be wrong
• The system prompt changes every call — prompt caching (Level 5) has no effect; skip straight to Level 1–3

The Reduction Hierarchy

Apply in order. Stop when the target is met. Do not apply all steps preemptively.

Level 1 — Trim the System Prompt

Audit every sentence:

• Run the prompt without each instruction — does output quality change?
• Remove any instruction the model follows without it
• Remove few-shot examples that duplicate knowledge the model already has
• Replace verbose explanations with single directives

Target: system prompt under 500 tokens for most tasks. Measure token count precisely:

import tiktoken
enc = tiktoken.get_encoding("cl100k_base")
print(len(enc.encode(system_prompt)))

Level 2 — Reduce Retrieved Context

If using RAG, before any other change:

1. Reduce top_k by 1 and run eval-before-ship — recall often holds at lower top_k
2. Add or tighten a similarity threshold (filter chunks below 0.75 cosine similarity)
3. Set a hard context budget: MAX_CONTEXT_TOKENS = context_window × 0.4

Level 3 — Compress Retrieved Content

Before injecting:

• Strip document headers, footers, and repeated boilerplate
• Normalise whitespace and remove formatting artefacts
• Convert verbose JSON to compact tabular format
• For chunks over 600 tokens: LLM-summarise to 1/3 length with a fast model

Level 4 — Split the Pipeline

Move expensive operations to a cheaper model:

Before: Frontier model does everything
After:  Fast/cheap model → classify, extract, summarise
        Frontier model → final answer generation only

This typically cuts costs 5–10× with minimal quality loss. The frontier model only sees pre-processed, high-signal input.

Level 5 — Cache Stable Context

If system prompt or retrieval context is identical across many calls:

• Enable prompt caching (Anthropic Beta, OpenAI prompt caching)
• Cache embedding results for repeated queries (TTL matched to data freshness)

Prompt caching saves 50–90% on the cached portion. It is the highest-leverage optimization when the system prompt is long and stable.

Measuring Before and After

Run this before applying any reduction, and after:

Token count:  <input tokens>, <output tokens>, <total>
Latency:      p50 = Xs, p95 = Ys
Cost/1k:      $Z
Quality:      <eval suite pass rate: A%>

A reduction that saves 30% cost but drops quality 10% past threshold is not an optimization — it is a regression.

Rationalization Red Flags

These thoughts mean the prompt has not been measured — stop:

• "I already have a short prompt" — count the tokens; "short" is not a token count
• "Reducing context will hurt quality" — you don't know until you measure with eval-before-ship
• "It costs too little to matter now" — $0.05/call is $1,500/month at 1k calls/day; growth events don't announce themselves; run the 10× projection from ai-cost-audit before calling it negligible
• "Prompt caching is complex to set up" — it is a one-line header change; do Level 5 before assuming it's hard
• "The frontier model is only marginally more expensive" — at 10k calls/day, "marginally" compounds

Completion Statement Format

When context-optimization is satisfied, state it like this:

Context optimized.
Reductions applied: <list of levels: e.g., L1 (trim), L2 (top_k 10→5), L5 (caching)>

Before:
  Input tokens: N (system: A, context: B, user: C)
  Latency p95: Xs
  Cost/1k: $Y
  Quality: Z% — evals/<feature>/results-<date-before>.md

After:
  Input tokens: N' (system: A', context: B', user: C')
  Latency p95: X's
  Cost/1k: $Y'
  Quality: Z'% — evals/<feature>/results-<date-after>.md ✓ (above threshold)

Delta: -X% cost, -Ys latency, quality delta: Z'% - Z% = ±Xpp

Quality must be verified with eval-before-ship — not eyeballed.

Why This Matters

LLM API costs compound with volume. A pipeline that costs $0.05/call at 1k calls/day costs $1,800/month. The same pipeline at 10k calls/day costs $18,000/month. Optimization at 1k is a 2-hour project. Optimization at 10k is a crisis sprint.