SAYA AI — an agentic
copilot for your data.
SAYA AI reads your schema, runs read-only probes, and returns Markdown reports with runnable SQL. It works with Claude, OpenAI, Gemini, or runs entirely local on Ollama.
Three modes
Every mode calls real tools to gather evidence before answering — SAYA doesn't guess.
Ask SAYA anything about your schema. It calls tools, gathers context, and answers in Markdown with runnable SQL blocks.
Run one of 6 curated profiling skills. SAYA gathers real statistics, authors profiling SQL, and returns a Markdown report plus runnable SQL.
Give SAYA an analysis goal. It plans the steps, runs SQL blocks, and streams results back. You control how much autonomy it has.
The tool catalog
SAYA doesn't guess — it calls tools to gather real evidence before answering.
Search tables and columns by business meaning.
Execute a read-only query to probe real data.
Fetch the CREATE TABLE statement.
Row count and column metadata.
Save durable notes about table/column business meaning.
Ask a clarifying question when more context is needed.
Stop gracefully when a question is unanswerable.
Search bundled SQL reference snippets.
Search your own saved SQL and markdown blocks.
Six profiling skills
Curated prompts instruct SAYA to gather real statistics, author profiling SQL, and return both a Markdown report and runnable SQL blocks.
profile_statsNull percentages, distinct counts, min/max, top values, and type validation.
quality_scoringPrimary key uniqueness, foreign key integrity, NOT-NULL compliance, and completeness scores.
anomaly_detectionNumeric outliers, rare categorical values, and unexpected NULL bursts.
schema_docsHuman-readable column docs with inferred semantic types and PII candidates.
proactive_insightsTop data issues ranked by severity with diagnostic SQL queries.
relationship_discoveryFunctional dependencies, key candidates, and recommended joins.
SAYA automatically adjusts its profiling approach for large tables to keep things fast.
Providers
Bring your own model. Cloud providers with native tool-calling, or run entirely on-device with Ollama.
| Provider | Mode | Detail |
|---|---|---|
| Claude (Anthropic) | Cloud | Cloud-based inference with tool-calling support |
| OpenAI / compatible | Cloud | OpenAI-compatible endpoints with tool support |
| Gemini (Google) | Cloud | Cloud-based inference with tool-calling support |
| Ollama | Local | Fully on-device inference — no network required |
| OpenAI-compatible local | Local | Connect any local OpenAI-compatible endpoint |
Safety & guardrails
- Read-only Guardrails — Every AI-authored query is validated to ensure only SELECT, SHOW, and EXPLAIN are permitted. DROP, DELETE, and TRUNCATE are blocked before execution — per database dialect.
- PII Scanner — Query results are scanned for sensitive data patterns including SSNs, credit cards, emails, phone numbers, and IBANs. Redacted previews keep sensitive data out of AI output.
- Enterprise Privacy Mode — When privacy mode is enabled, the AI cannot execute queries against your data. It reasons from schema metadata only and authors SQL for you to run — no data leaves your machine.
- Audit Log — Every AI run is logged locally with mode, target, tool calls, guardrail blocks, rows scanned, and outcome — available for your review anytime.
- Repetition Guard — A smart budget system prevents the AI from spinning on near-identical queries. Frequency and consecutive-repeat limits abort unproductive loops automatically.
- Iteration Cap + Cancellation — A hard iteration cap prevents runaway runs. You can cancel any AI operation mid-flight at any time.