arize-prompt-optimization

# Arize Prompt Optimization Skill

## Concepts

### Where Prompts Live in Trace Data

LLM applications emit spans following OpenInference semantic conventions. Prompts are stored in different span attributes depending on the span kind and instrumentation:

| Column | What it contains | When to use |
|--------|-----------------|-------------|
| `attributes.llm.input_messages` | Structured chat messages (system, user, assistant, tool) in role-based format | **Primary source** for chat-based LLM prompts |
| `attributes.llm.input_messages.roles` | Array of roles: `system`, `user`, `assistant`, `tool` | Extract individual message roles |
| `attributes.llm.input_messages.contents` | Array of message content strings | Extract message text |
| `attributes.input.value` | Serialized prompt or user question (generic, all span kinds) | Fallback when structured messages are not available |
| `attributes.llm.prompt_template.template` | Template with `{variable}` placeholders (e.g., `"Answer {question} using {context}"`) | When the app uses prompt templates |
| `attributes.llm.prompt_template.variables` | Template variable values (JSON object) | See what values were substituted into the template |
| `attributes.output.value` | Model response text | See what the LLM produced |
| `attributes.llm.output_messages` | Structured model output (including tool calls) | Inspect tool-calling responses |

### Finding Prompts by Span Kind

- **LLM span** (`attributes.openinference.span.kind = 'LLM'`): Check `attributes.llm.input_messages` for structured chat messages, OR `attributes.input.value` for a serialized prompt. Check `attributes.llm.prompt_template.template` for the template.
- **Chain/Agent span**: `attributes.input.value` contains the user's question. The actual LLM prompt lives on **child LLM spans** -- navigate down the trace tree.
- **Tool span**: `attributes.input.value` has tool input, `attributes.output.value` has tool result. Not typically where prompts live.

### Performance Signal Columns

These columns carry the feedback data used for optimization:

| Column pattern | Source | What it tells you |
|---------------|--------|-------------------|
| `annotation.<name>.label` | Human reviewers | Categorical grade (e.g., `correct`, `incorrect`, `partial`) |
| `annotation.<name>.score` | Human reviewers | Numeric quality score (e.g., 0.0 - 1.0) |
| `annotation.<name>.text` | Human reviewers | Freeform explanation of the grade |
| `eval.<name>.label` | LLM-as-judge evals | Automated categorical assessment |
| `eval.<name>.score` | LLM-as-judge evals | Automated numeric score |
| `eval.<name>.explanation` | LLM-as-judge evals | Why the eval gave that score -- **most valuable for optimization** |
| `attributes.input.value` | Trace data | What went into the LLM |
| `attributes.output.value` | Trace data | What the LLM produced |
| `{experiment_name}.output` | Experiment runs | Output from a specific experiment |

## Prerequisites

Proceed directly with the task — run the `ax` command you need. Do NOT check versions, env vars, or profiles upfront.

If an `ax` command fails, troubleshoot based on the error:
- `command not found` or version error → see references/ax-setup.md
- `401 Unauthorized` / missing API key → run `ax profiles show` to inspect the current profile. If the profile is missing or the API key is wrong: check `.env` for `ARIZE_API_KEY` and use it to create/update the profile via references/ax-profiles.md. If `.env` has no key either, ask the user for their Arize API key (https://app.arize.com/admin > API Keys)
- Space ID unknown → check `.env` for `ARIZE_SPACE_ID`, or run `ax spaces list -o json`, or ask the user
- Project unclear → check `.env` for `ARIZE_DEFAULT_PROJECT`, or ask, or run `ax projects list -o json --limit 100` and present as selectable options
- LLM provider call fails (missing OPENAI_API_KEY / ANTHROPIC_API_KEY) → check `.env`, load if present, otherwise ask the user

## Phase 1: Extract the Current Prompt

### Find LLM spans containing prompts

```bash
# List LLM spans (where prompts live)
ax spans list PROJECT_ID --filter "attributes.openinference.span.kind = 'LLM'" --limit 10

# Filter by model
ax spans list PROJECT_ID --filter "attributes.llm.model_name = 'gpt-4o'" --limit 10

# Filter by span name (e.g., a specific LLM call)
ax spans list PROJECT_ID --filter "name = 'ChatCompletion'" --limit 10
```

### Export a trace to inspect prompt structure

```bash
# Export all spans in a trace
ax spans export --trace-id TRACE_ID --project PROJECT_ID

# Export a single span
ax spans export --span-id SPAN_ID --project PROJECT_ID
```

### Extract prompts from exported JSON

```bash
# Extract structured chat messages (system + user + assistant)
jq '.[0] | {
  messages: .attributes.llm.input_messages,
  model: .attributes.llm.model_name
}' trace_*/spans.json

# Extract the system prompt specifically
jq '[.[] | select(.attributes.llm.input_messages.roles[]? == "system")] | .[0].attributes.llm.input_messages' trace_*/spans.json

# Extract prompt template and variables
jq '.[0].attributes.llm.prompt_template' trace_*/spans.json

# Extract from input.value (fallback for non-structured prompts)
jq '.[0].attributes.input.value' trace_*/spans.json
```

### Reconstruct the prompt as messages

Once you have the span data, reconstruct the prompt as a messages array:

```json
[
  {"role": "system", "content": "You are a helpful assistant that..."},
  {"role": "user", "content": "Given {input}, answer the question: {question}"}
]
```

If the span has `attributes.llm.prompt_template.template`, the prompt uses variables. Preserve these placeholders (`{variable}` or `{{variable}}`) -- they are substituted at runtime.

## Phase 2: Gather Performance Data

### From traces (production feedback)

```bash
# Find error spans -- these indicate prompt failures
ax spans list PROJECT_ID \
  --filter "status_code = 'ERROR' AND attributes.openinference.span.kind = 'LLM'" \
  --limit 20

# Find spans with low eval scores
ax spans list PROJECT_ID \
  --filter "annotation.correctness.label = 'incorrect'" \
  --limit 20

# Find spans with high latency (may indicate overly complex prompts)
ax spans list PROJECT_ID \
  --filter "attributes.openinference.span.kind = 'LLM' AND latency_ms > 10000" \
  --limit 20

# Export error traces for detailed inspection
ax spans export --trace-id TRACE_ID --project PROJECT_ID
```

### From datasets and experiments

```bash
# Export a dataset (ground truth examples)
ax datasets export DATASET_ID
# -> dataset_*/examples.json

# Export experiment results (what the LLM produced)
ax experiments export EXPERIMENT_ID
# -> experiment_*/runs.json
```

### Merge dataset + experiment for analysis

Join the two files by `example_id` to see inputs alongside outputs and evaluations:

```bash
# Count examples and runs
jq 'length' dataset_*/examples.json
jq 'length' experiment_*/runs.json

# View a single joined record
jq -s '
  .[0] as $dataset |
  .[1][0] as $run |
  ($dataset[] | select(.id == $run.example_id)) as $example |
  {
    input: $example,
    output: $run.output,
    evaluations: $run.evaluations
  }
' dataset_*/examples.json experiment_*/runs.json

# Find failed examples (where eval score < threshold)
jq '[.[] | select(.evaluations.correctness.score < 0.5)]' experiment_*/runs.json
```

### Identify what to optimize

Look for patterns across failures:

1. **Compare outputs to ground truth**: Where does the LLM output differ from expected?
2. **Read eval explanations**: `eval.*.explanation` tells you WHY something failed
3. **Check annotation text**: Human feedback describes specific issues
4. **Look for verbosity mismatches**: If outputs are too long/short vs ground truth
5. **Check format compliance**: Are outputs in the expected format?

## Phase 3: Optimize the Prompt

### The Optimization Meta-Prompt

Use this template to generate an improved version of the prompt. Fill in the three placeholders and send it to your LLM (GPT-4o, Claude, etc.):

````
You are an expert in prompt optimization. Given the original baseline prompt
and the associated performance data (inputs, outputs, evaluation labels, and
explanations), generate a revised version that improves results.

ORIGINAL BASELINE PROMPT
========================

{PASTE_ORIGINAL_PROMPT_HERE}

========================

PERFORMANCE DATA
================

The following records show how the current prompt performed. Each record
includes the input, the LLM output, and evaluation feedback:

{PASTE_RECORDS_HERE}

================

HOW TO USE THIS DATA

1. Compare outputs: Look at what the LLM generated vs what was expected
2. Review eval scores: Check which examples scored poorly and why
3. Examine annotations: Human feedback shows what worked and what didn't
4. Identify patterns: Look for common issues across multiple examples
5. Focus on failures: The rows where the output DIFFERS from the expected
   value are the ones that need fixing

ALIGNMENT STRATEGY

- If outputs have extra text or reasoning not present in the ground truth,
  remove instructions that encourage explanation or verbose reasoning
- If outputs are missing information, add instructions to include it
- If outputs are in the wrong format, add explicit format instructions
- Focus on the rows where the output differs from the target -- these are
  the failures to fix

RULES

Maintain Structure:
- Use the same template variables as the current prompt ({var} or {{var}})
- Don't change sections that are already working
- Preserve the exact return format instructions from the original prompt

Avoid Overfitting:
- DO NOT copy examples verbatim into the prompt
- DO NOT quote specific test data outputs exactly
- INSTEAD: Extract the ESSENCE of what makes good vs bad outputs
- INSTEAD: Add general guidelines and principles
- INSTEAD: If adding few-shot examples, create SYNTHETIC examples that
  demonstrate the principle, not real data from above

Goal: Create a prompt that generalizes well to new inputs, not one that
memorizes the test data.

OUTPUT FORMAT

Return the revised prompt as a JSON array of messages:

[
  {"role": "system", "content": "..."},
  {"role": "user", "content": "..."}
]

Also provide a brief reasoning section (bulleted list) explaining:
- What problems you found
- How the revised prompt addresses each one
````

### Preparing the performance data

Format the records as a JSON array before pasting into the template:

```bash
# From dataset + experiment: join and select relevant columns
jq -s '
  .[0] as $ds |
  [.[1][] | . as $run |
    ($ds[] | select(.id == $run.example_id)) as $ex |
    {
      input: $ex.input,
      expected: $ex.expected_output,
      actual_output: $run.output,
      eval_score: $run.evaluations.correctness.score,
      eval_label: $run.evaluations.correctness.label,
      eval_explanation: $run.evaluations.correctness.explanation
    }
  ]
' dataset_*/examples.json experiment_*/runs.json

# From exported spans: extract input/output pairs with annotations
jq '[.[] | select(.attributes.openinference.span.kind == "LLM") | {
  input: .attributes.input.value,
  output: .attributes.output.value,
  status: .status_code,
  model: .attributes.llm.model_name
}]' trace_*/spans.json
```

### Applying the revised prompt

After the LLM returns the revised messages array:

1. Compare the original and revised prompts side by side
2. Verify all template variables are preserved
3. Check that format instructions are intact
4. Test on a few examples before full deployment

## Phase 4: Iterate

### The optimization loop

```
1. Extract prompt    -> Phase 1 (once)
2. Run experiment    -> ax experiments create ...
3. Export results    -> ax experiments export EXPERIMENT_ID
4. Analyze failures  -> jq to find low scores
5. Run meta-prompt   -> Phase 3 with new failure data
6. Apply revised prompt
7. Repeat from step 2
```

### Measure improvement

```bash
# Compare scores across experiments
# Experiment A (baseline)
jq '[.[] | .evaluations.correctness.score] | add / length' experiment_a/runs.json

# Experiment B (optimized)
jq '[.[] | .evaluations.correctness.score] | add / length' experiment_b/runs.json

# Find examples that flipped from fail to pass
jq -s '
  [.[0][] | select(.evaluations.correctness.label == "incorrect")] as $fails |
  [.[1][] | select(.evaluations.correctness.label == "correct") |
    select(.example_id as $id | $fails | any(.example_id == $id))
  ] | length
' experiment_a/runs.json experiment_b/runs.json
```

### A/B compare two prompts

1. Create two experiments against the same dataset, each using a different prompt version
2. Export both: `ax experiments export EXP_A` and `ax experiments export EXP_B`
3. Compare average scores, failure rates, and specific example flips
4. Check for regressions -- examples that passed with prompt A but fail with prompt B

## Prompt Engineering Best Practices

Apply these when writing or revising prompts:

| Technique | When to apply | Example |
|-----------|--------------|---------|
| Clear, detailed instructions | Output is vague or off-topic | "Classify the sentiment as exactly one of: positive, negative, neutral" |
| Instructions at the beginning | Model ignores later instructions | Put the task description before examples |
| Step-by-step breakdowns | Complex multi-step processes | "First extract entities, then classify each, then summarize" |
| Specific personas | Need consistent style/tone | "You are a senior financial analyst writing for institutional investors" |
| Delimiter tokens | Sections blend together | Use `---`, `###`, or XML tags to separate input from instructions |
| Few-shot examples | Output format needs clarification | Show 2-3 synthetic input/output pairs |
| Output length specifications | Responses are too long or short | "Respond in exactly 2-3 sentences" |
| Reasoning instructions | Accuracy is critical | "Think step by step before answering" |
| "I don't know" guidelines | Hallucination is a risk | "If the answer is not in the provided context, say 'I don't have enough information'" |

### Variable preservation

When optimizing prompts that use template variables:

- **Single braces** (`{variable}`): Python f-string / Jinja style. Most common in Arize.
- **Double braces** (`{{variable}}`): Mustache style. Used when the framework requires it.
- Never add or remove variable placeholders during optimization
- Never rename variables -- the runtime substitution depends on exact names
- If adding few-shot examples, use literal values, not variable placeholders

## Workflows

### Optimize a prompt from a failing trace

1. Find failing traces:
   ```bash
   ax traces list PROJECT_ID --filter "status_code = 'ERROR'" --limit 5
   ```
2. Export the trace:
   ```bash
   ax spans export --trace-id TRACE_ID --project PROJECT_ID
   ```
3. Extract the prompt from the LLM span:
   ```bash
   jq '[.[] | select(.attributes.openinference.span.kind == "LLM")][0] | {
     messages: .attributes.llm.input_messages,
     template: .attributes.llm.prompt_template,
     output: .attributes.output.value,
     error: .attributes.exception.message
   }' trace_*/spans.json
   ```
4. Identify what failed from the error message or output
5. Fill in the optimization meta-prompt (Phase 3) with the prompt and error context
6. Apply the revised prompt

### Optimize using a dataset and experiment

1. Find the dataset and experiment:
   ```bash
   ax datasets list
   ax experiments list --dataset-id DATASET_ID
   ```
2. Export both:
   ```bash
   ax datasets export DATASET_ID
   ax experiments export EXPERIMENT_ID
   ```
3. Prepare the joined data for the meta-prompt
4. Run the optimization meta-prompt
5. Create a new experiment with the revised prompt to measure improvement

### Debug a prompt that produces wrong format

1. Export spans where the output format is wrong:
   ```bash
   ax spans list PROJECT_ID \
     --filter "attributes.openinference.span.kind = 'LLM' AND annotation.format.label = 'incorrect'" \
     --limit 10 -o json > bad_format.json
   ```
2. Look at what the LLM is producing vs what was expected
3. Add explicit format instructions to the prompt (JSON schema, examples, delimiters)
4. Common fix: add a few-shot example showing the exact desired output format

### Reduce hallucination in a RAG prompt

1. Find traces where the model hallucinated:
   ```bash
   ax spans list PROJECT_ID \
     --filter "annotation.faithfulness.label = 'unfaithful'" \
     --limit 20
   ```
2. Export and inspect the retriever + LLM spans together:
   ```bash
   ax spans export --trace-id TRACE_ID --project PROJECT_ID
   jq '[.[] | {kind: .attributes.openinference.span.kind, name, input: .attributes.input.value, output: .attributes.output.value}]' trace_*/spans.json
   ```
3. Check if the retrieved context actually contained the answer
4. Add grounding instructions to the system prompt: "Only use information from the provided context. If the answer is not in the context, say so."

## Troubleshooting

| Problem | Solution |
|---------|----------|
| `ax: command not found` | See references/ax-setup.md |
| `No profile found` | No profile is configured. See references/ax-profiles.md to create one. |
| No `input_messages` on span | Check span kind -- Chain/Agent spans store prompts on child LLM spans, not on themselves |
| Prompt template is `null` | Not all instrumentations emit `prompt_template`. Use `input_messages` or `input.value` instead |
| Variables lost after optimization | Verify the revised prompt preserves all `{var}` placeholders from the original |
| Optimization makes things worse | Check for overfitting -- the meta-prompt may have memorized test data. Ensure few-shot examples are synthetic |
| No eval/annotation columns | Run evaluations first (via Arize UI or SDK), then re-export |
| Experiment output column not found | The column name is `{experiment_name}.output` -- check exact experiment name via `ax experiments get` |
| `jq` errors on span JSON | Ensure you're targeting the correct file path (e.g., `trace_*/spans.json`) |