Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.aevyra.ai/llms.txt

Use this file to discover all available pages before exploring further.

After optimization, reflex produces a detailed analysis that explains what happened and teaches prompt engineering principles. This page walks through each section.

​
Score trajectory

Trajectory : 0.350 → 0.450 → 0.520 → 0.600 → 0.650 → 0.720 → 0.780 → 0.850 → 0.880
The trajectory shows every iteration’s score. Reflex analyzes the shape:
  • Steady climb — consistent improvement across iterations
  • Plateau — scores flatten, suggesting diminishing returns from the current approach
  • Over-optimization — scores peak then regress (model may be overfitting to a pattern)
  • Gap closed — how much of the remaining gap (to 1.0) was closed
If the result didn’t converge, reflex suggests next steps: trying a different strategy, adding more data, or adjusting the threshold.

​
Strategy breakdown

When using auto mode, reflex shows what each phase contributed: 
Strategy breakdown:
  Phase 1 — structural    : 0.350 → 0.520 (+0.170)
  Phase 2 — iterative     : 0.520 → 0.780 (+0.260)
  Phase 3 — fewshot       : 0.780 → 0.880 (+0.100)
Each phase also includes an educational lesson explaining why that technique helped (or didn’t):
  • Structural helped — “Structure matters: reorganizing how instructions are presented can dramatically improve model comprehension.”
  • Iterative helped — “Specificity matters: models follow precise, explicit instructions better than vague ones.”
  • Fewshot helped — “Examples matter: showing the model what good output looks like is one of the most reliable ways to improve quality.”
  • Phase hurt performance — the analysis explains what went wrong and what to avoid

​
Prompt diff

What changed in the prompt:
  - Much longer (5 → 47 words). The model needed more detailed instructions.
  - Added: markdown headers for clear sections, bold emphasis on key
    instructions, XML tags for structural clarity, explicit constraints
    on what to avoid
Reflex compares the original and optimized prompts, highlighting:
  • Length changes and what they mean
  • New structural features (headers, bullets, XML tags, examples)
  • Added constraints or format specifications

​
Before / after example

BEFORE / AFTER EXAMPLE (most-improved sample):
  Input:  Summarize photosynthesis.
  BEFORE (score: 0.25):
    Plants use light.
  AFTER (score: 0.75):
    Plants convert sunlight into chemical energy using chlorophyll...
  Score change: 0.25 → 0.75 (+0.50)
Reflex picks the sample with the largest score improvement and shows the concrete difference the optimized prompt made.

​
Programmatic access

All analysis data is available programmatically: 
result = optimizer.run("You are a helpful assistant.")

# Full formatted summary (includes all sections above)
print(result.summary())

# Raw data
result.score_trajectory       # [0.35, 0.45, ...]
result.improvement            # 0.58
result.improvement_pct        # 193.3
result.phase_history          # [{"phase": 1, "axis": "structural", ...}]

# Serialize
result.to_json("results.json")
result.save_best_prompt("best_prompt.md")