Full Config Reference¶
config/config.yaml is the single source of truth, loaded once at startup from a fixed path (<project root>/config/config.yaml). Every string value supports ${ENV_VAR} (required) or ${ENV_VAR:default} (optional) interpolation.
Minimal Configuration¶
Only llm and database are required:
llm:
provider: "anthropic"
anthropic:
api_key: "${ANTHROPIC_API_KEY}"
model: "claude-sonnet-4-20250514"
database:
provider: "trino"
trino:
host: "${TRINO_HOST}"
port: 443
user: "${TRINO_USER}"
password: "${TRINO_PASSWORD}"
catalog: "hive"
schema: "default"
http_scheme: "https"
app¶
llm¶
llm:
provider: "anthropic" # anthropic | gemini | vertexai
anthropic:
api_key: "${ANTHROPIC_API_KEY}"
model: "claude-sonnet-4-20250514"
max_tokens: 4096
gemini:
api_key: "${GEMINI_API_KEY}"
model: "gemini-pro"
max_tokens: 1000
vertexai:
project_id: "${VERTEXAI_PROJECT_ID}"
location: "${VERTEXAI_LOCATION:europe-west1}"
model: "gemini-2.5-pro"
credentials_path: "${GOOGLE_APPLICATION_CREDENTIALS}"
# or inline for containers: credentials_json: "${VERTEXAI_CREDENTIALS_JSON}"
model_routing:
intent_detection: "gemini-2.5-flash"
chat: "gemini-2.5-flash"
sql: "gemini-2.5-pro"
insights: "gemini-2.5-flash"
ops_insights: "gemini-2.5-flash"
model_routing (top-level, works across providers) routes each task type to a different model/provider to balance cost and accuracy — see LLM Providers.
database¶
database:
provider: "trino" # trino | postgresql | mysql | bigquery
trino:
host: "${TRINO_HOST}"
port: 443
user: "${TRINO_USER}"
password: "${TRINO_PASSWORD}"
catalog: "${TRINO_CATALOG:hive}"
schema: "${TRINO_SCHEMA:default}"
http_scheme: "https"
timeout: 300 # optional, seconds — server-side query cancellation
postgresql:
host: "${POSTGRES_HOST}"
port: 5432
database: "${POSTGRES_DB}"
user: "${POSTGRES_USER}"
password: "${POSTGRES_PASSWORD}"
schema: "${POSTGRES_SCHEMA:public}"
min_connections: 1
max_connections: 10
timeout: 300 # optional, seconds — server-side query cancellation
mysql:
host: "${MYSQL_HOST}"
port: 3306
database: "${MYSQL_DB}"
user: "${MYSQL_USER}"
password: "${MYSQL_PASSWORD}"
autocommit: true
timeout: 300 # optional, seconds — server-side query cancellation
bigquery:
project_id: "${BQ_PROJECT_ID}"
dataset: ""
location: "US"
credentials_path: "${GOOGLE_APPLICATION_CREDENTIALS}"
# or: credentials_json: ""
timeout: 300
max_results: 10000
See Database Providers for field-by-field notes, including how each provider cancels a timed-out query.
mcp¶
mcp:
sse:
enabled: true # exposes GET /mcp/sse + POST /mcp/messages/
# Notify when an AI/MCP-generated query looks sensitive. Fires on either
# trigger — a risk_score above the threshold, or the query referencing a
# column matching an enabled MCP Masking field_pattern. Disabled by
# default; every check is best-effort and never blocks the query.
security_alerts:
enabled: false
risk_threshold: 70 # 0-100, see sql_validator.risk_score()
apply_to: ["mcp"] # which widget_type values to alert on
notify_emails: [] # requires email.enabled: true; always logs regardless
See MCP Network Setup. Local (stdio) MCP mode needs no config here — it authenticates via SQLATTE_TOKEN/env vars passed to sqlatte_mcp_server.py directly.
admin¶
Fails closed if password is empty — see Admin Panel Authentication.
ldap¶
ldap:
enabled: false
server: "${LDAP_SERVER:ldaps://ldap.example.com:636}"
use_ssl: true
connect_timeout: 5
# Direct bind (simplest, no service account):
user_dn_template: "uid={username},ou=people,dc=example,dc=com"
# Or search+bind with a service account:
# bind_dn: "${LDAP_BIND_DN}"
# bind_password: "${LDAP_BIND_PASSWORD}"
# base_dn: "${LDAP_BASE_DN}"
# user_search_filter: "(uid={username})" # or "(sAMAccountName={username})" for AD
# Optional: restrict login to one LDAP group
# required_group_dn: "cn=sqlatte-users,ou=groups,dc=example,dc=com"
# group_member_attr: "member" # "member" (OpenLDAP) | "memberOf" (AD) | "memberUid" (POSIX)
Disabled by default. When enabled, used by admin login (tried first, falling back to admin.username/password if the directory is unreachable or the bind fails) and, optionally, the Assistant login gate (plugins.assistant_login, below). Also editable from the Admin Panel's LDAP / SSO tab, with optional persistence to config_db. See Admin Panel Authentication.
plugins.assistant_login¶
plugins:
assistant_login:
enabled: false
ttl_hours: 8 # session lifetime after a successful LDAP login
Optional LDAP login gate in front of the SQLatte Assistant (the legacy /query widget), which otherwise has no login at all. Requires ldap.enabled: true. On success this is an identity check only — the issued session still uses the server's own database config, not per-user credentials (unlike plugins.auth, below).
config_db¶
config_db:
enabled: false
type: postgresql
postgresql:
host: "${CONFIG_DB_HOST:localhost}"
database: "${CONFIG_DB_NAME:sqlatte_config}"
user: "${CONFIG_DB_USER:sqlatte}"
password: "${CONFIG_DB_PASSWORD}"
# Generate with: python3 -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
encryption_key: "${CONFIG_DB_ENCRYPTION_KEY}"
When enabled, runtime config changes, API tokens, mask rules, and audit logs persist to PostgreSQL instead of process memory — required for anything beyond single-instance/dev use. encryption_key encrypts stored credentials at rest.
plugins.auth¶
plugins:
auth:
enabled: false
session_ttl_minutes: 480
max_workers: 40
db_provider: "trino"
db_host: "${TRINO_HOST}"
db_port: 443
allowed_catalogs:
- name: "hadoop"
allowed_schemas: ["hive"]
allowed_db_types: ["trino"]
# Trino only, on by default: reject generated SQL that references a
# catalog other than the one the session/token is scoped to. No effect
# without allowed_catalogs configured or on non-Trino providers.
enforce_catalog_lock: true
# Trino only, off by default: catalog-less tokens scoped by
# allowed_catalogs instead of one fixed catalog — enables the MCP
# discover_tables tool and the /auth/discover* endpoints. See
# MCP Overview and API Tokens & Admin Auth.
enable_discovery_tokens: false
auto_session:
enabled: false
ttl_hours: 1
label: "auto-session"
# allowed_origins: []
Multi-tenant per-user database connections. See Security Overview and Discovery Tokens.
analytics¶
analytics:
enabled: false
backend: "postgresql"
postgresql:
host: "${ANALYTICS_DB_HOST:localhost}"
port: 5432
database: "${ANALYTICS_DB_NAME:sqlatte_analytics}"
user: "${ANALYTICS_DB_USER:sqlatte}"
password: "${ANALYTICS_DB_PASSWORD}"
See Analytics Setup.
scheduler / email / export¶
scheduler:
enabled: false
timezone: "UTC"
max_concurrent_jobs: 10
job_timeout_seconds: 300
keep_history_days: 30
max_executions_per_schedule: 100
email:
enabled: false
provider: "smtp"
smtp:
host: "${SMTP_HOST}"
port: 587
user: "${SMTP_USER:}"
password: "${SMTP_PASSWORD:}"
from_email: "${SMTP_FROM_EMAIL}"
from_name: "SQLatte Reports"
templates:
success_subject: "✅ {{schedule_name}} - {{date}}"
failure_subject: "❌ Failed: {{schedule_name}}"
max_emails_per_day: 1000
max_recipients_per_email: 10
export:
formats: [csv, excel, html]
max_rows: 10000
max_file_size_mb: 25
filename_template: "{{schedule_name}}_{{date}}_{{time}}.{{format}}"
See Scheduled Queries and Email Settings.
insights¶
insights:
enabled: true
mode: hybrid # llm_only | statistical_only | hybrid
max_insights: 3
include_statistical: true
ops_agent / alarms / jira¶
ops_agent:
enabled: false
ai_insights: true
ai_insights_max: 5
provider: "bigquery_ops"
config:
projects:
- project_id: "${BQ_OPS_PROJECT_ID}"
region: "${BQ_OPS_REGION:europe-west1}"
credentials_path: "${BQ_OPS_CREDENTIALS_PATH}"
allowed_operations:
- get_expensive_queries
- forecast_monthly_costs
- find_public_datasets
- check_slot_saturation
# ... see Ops Console page for the full list
alarms:
enabled: false
storage_path: "config/alarms.json"
default_recipients: ["${ALARM_DEFAULT_RECIPIENT}"]
jira:
enabled: false
url: "${JIRA_URL}"
email: "${JIRA_EMAIL}"
api_token: "${JIRA_API_TOKEN}"
project_key: "${JIRA_PROJECT_KEY:DATA}"
issue_type: "Bug"
priority: "High"
See BigQuery Ops Console for the full allowed_operations list and alarm behavior.
rate_limiting¶
rate_limiting:
enabled: false
strategy: sliding_window # sliding_window | token_bucket | fixed_window
key_type: session_id # session_id | ip | user_id
requests_per_window: 30
window_seconds: 60
protected_paths: ["/query", "/auth/query"]
exclude_paths: ["/health", "/static", "/admin"]
# Optional per-path limits, taking priority over the defaults above for
# matching paths — a path here doesn't need to also be in protected_paths.
path_overrides:
"/auth/query":
requests_per_window: 10
window_seconds: 60
"/auth/discover":
requests_per_window: 20
window_seconds: 60
query¶
query:
default_limit: 100
max_limit: 1000
timeout: 300
require_session: true # 401 unauthenticated callers on the legacy /query endpoint
# If false, the assistant generates SQL but does not run it — only affects
# the interactive assistant (/query, /auth/query); scheduled queries and
# MCP tool calls always execute regardless of this flag.
execute_generated_sql: true
# Ask the user to confirm (instead of just displaying or silently running
# the generated SQL) before executing it. See POST /query/confirm-execute
# and POST /auth/query/confirm in the API Reference.
confirm_before_execute: false
database:
bigquery:
price_per_tb_usd: 6.25 # optional — adds a $ estimate to the confirm prompt (BigQuery only)
confirm_before_execute applies to the interactive Assistant and Authenticated Agent widgets only — programmatic access (MCP tools, API tokens) always executes immediately. price_per_tb_usd lives alongside the rest of database.bigquery (see Database Providers), not under query.
cors¶
Restrict allow_origins to specific domains before embedding widgets in production.
ui.sections¶
ui:
sections:
assistant: true
demo: true
analytics: false
schedules: true
dashboards: true
bigquery_ops: true # or "bigquery-ops" — both spellings accepted
admin: true
tokens: true
Toggles which tabs appear in the standalone dashboard interface, independent of whether the underlying feature is enabled. home is always shown and isn't a toggleable key. Any key omitted from ui.sections defaults to true.
prompts¶
Four prompts — intent_detection, barista_personality, sql_generation, insights_generation (plus ops_insights_generation when the Ops Console is on) — are defined here as the defaults, and editable at runtime from the admin panel. See Runtime Prompts.