> ## Documentation Index
> Fetch the complete documentation index at: https://villagesql.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Summarization in MySQL

> How to summarize long text stored in MySQL using VillageSQL's ai_prompt() — generating summaries of support tickets, articles, and documents directly in SQL.

<Card title="VillageSQL is a drop-in replacement for MySQL with extensions." icon="database" href="/mysql-8.4/0.0.4/quickstart">
  All examples in this guide work on VillageSQL. Install Now →
</Card>

Summarization converts long text into a shorter version that captures the key points. Support tickets, customer feedback, articles, and log entries all benefit from a short summary field that lets you scan content quickly. MySQL has no summarization function. VillageSQL's `ai_prompt()` adds one.

## With VillageSQL: Summaries in SQL

```sql theme={null}
INSTALL EXTENSION vsql_ai;
```

A summarization prompt specifies the desired length and what to keep:

```sql theme={null}
SET @key = 'sk-ant-your-api-key';

ALTER TABLE support_tickets ADD COLUMN summary VARCHAR(500);

SET SESSION max_execution_time = 300000;

UPDATE support_tickets
SET summary = ai_prompt(
    'anthropic',
    'claude-haiku-4-5-20251001',
    @key,
    CONCAT(
        'Summarize this support ticket in one sentence. ',
        'Include the main issue and any resolution mentioned.\n\n',
        'Ticket: ', body
    )
)
WHERE summary IS NULL
LIMIT 100;
```

Query summaries like any other column — no AI call at read time:

```sql theme={null}
-- Recent tickets with AI summaries for quick triage
SELECT id, created_at, summary
FROM support_tickets
WHERE created_at > NOW() - INTERVAL 24 HOUR
ORDER BY created_at DESC;
```

### Structured summaries

For use cases where you need to extract specific fields, ask for JSON output:

```sql theme={null}
ALTER TABLE support_tickets ADD COLUMN ai_extract JSON;

UPDATE support_tickets
SET ai_extract = ai_prompt(
    'anthropic',
    'claude-haiku-4-5-20251001',
    @key,
    CONCAT(
        'Extract key information from this support ticket. ',
        'Return JSON with keys: issue (string), urgency (low/medium/high), ',
        'product_mentioned (string or null), resolution_needed (boolean). ',
        'Return only valid JSON.\n\n',
        'Ticket: ', body
    )
)
WHERE ai_extract IS NULL
LIMIT 100;

-- Triage: high-urgency tickets with no resolution
SELECT id, body,
       JSON_UNQUOTE(JSON_EXTRACT(ai_extract, '$.issue')) AS issue,
       JSON_UNQUOTE(JSON_EXTRACT(ai_extract, '$.product_mentioned')) AS product
FROM support_tickets
WHERE JSON_UNQUOTE(JSON_EXTRACT(ai_extract, '$.urgency')) = 'high'
ORDER BY created_at DESC;
```

### Summarizing multi-column content

Build richer context for the model by joining multiple fields:

```sql theme={null}
UPDATE articles
SET summary = ai_prompt(
    'anthropic',
    'claude-haiku-4-5-20251001',
    @key,
    CONCAT(
        'Write a two-sentence summary of this article for a preview card.\n\n',
        'Title: ', title, '\n',
        'Author: ', author_name, '\n',
        'Body: ', LEFT(body, 3000)   -- trim very long bodies
    )
)
WHERE summary IS NULL
LIMIT 50;
```

`LEFT(body, 3000)` avoids sending excessively long text to the model. Most AI models have a token limit; for very long documents, summarize the first few thousand characters or chunk the content.

## Length and Style Control

Prompt phrasing controls output length and tone:

| Goal              | Prompt phrasing                                     |
| :---------------- | :-------------------------------------------------- |
| One sentence      | "Summarize in one sentence."                        |
| Tweet-length      | "Summarize in 280 characters or fewer."             |
| Executive summary | "Write a 2–3 sentence executive summary."           |
| Bullet points     | "Summarize as 3 bullet points."                     |
| Action-oriented   | "What does this ticket ask us to do? One sentence." |

For summaries used in UI previews, specify a character limit to avoid overflow.

## Model Selection

Summarization is generally well-handled by faster, cheaper models. Use `claude-haiku-4-5-20251001` or `gpt-4o-mini` for bulk summarization. Step up to `claude-sonnet-4-5-20250929` when:

* The source text is long and complex
* You need structured JSON extraction with nuanced fields
* Accuracy of extracted facts matters more than cost

For setup and provider options, see [Connecting MySQL to AI APIs](/guides/ai-api-setup).

## Frequently Asked Questions

#### How long can the source text be?

Practical limit is a few thousand words per call. AI models have token limits (typically 100K–200K tokens for current models), but very long inputs cost more and take longer. For documents over \~5000 words, summarize the first section or chunk and summarize each chunk.

#### Can I use summaries in a full-text search index?

Yes — add a FULLTEXT index on the summary column after populating it. Summaries are often better targets for full-text search than raw content because they're shorter and more keyword-dense.

#### Will the same document always produce the same summary?

No — AI models are probabilistic. The same prompt can produce slightly different output on different calls. If consistency matters, generate the summary once and store it; don't regenerate on each read.

#### How do I summarize rows that have already been summarized?

Don't re-summarize unless the source content changed. Use `WHERE summary IS NULL` to only process new or updated rows. If you need to track changes, add a `summary_updated_at` column and compare it to the content's last-modified timestamp.

## Troubleshooting

| Problem                                              | Solution                                                    |
| :--------------------------------------------------- | :---------------------------------------------------------- |
| `FUNCTION ai_prompt does not exist`                  | Run `INSTALL EXTENSION vsql_ai`                             |
| Returns NULL                                         | Check API key, `max_execution_time`, and model name         |
| Summary too long or short                            | Tighten the prompt: specify a character or sentence limit   |
| Model repeats the source text instead of summarizing | Add "Do not copy sentences from the original" to the prompt |
| Structured JSON extraction has missing keys          | Add "If a field is not mentioned, use null" to the prompt   |

## See also

* [Running AI Models from MySQL Queries](/guides/ai-prompts-in-mysql) — the ai\_prompt() function this guide builds on
* [Sentiment Analysis on MySQL Data](/guides/sentiment-analysis) — another text analysis pattern with ai\_prompt()
* [Classifying Text in MySQL with AI](/guides/text-classification) — categorizing text using the same approach