Artificial intelligence

AI is optional. With no provider configured, nothing goes anywhere. basedb can talk to OpenAI, Anthropic and Mistral, with your own key — and to any server that speaks OpenAI’s API: Azure, a company gateway, a model you host yourself.

Configuring a provider

As long as no setting is saved in the interface, the API reads its environment:

BASEDB_AI_PROVIDER=mistral      # openai, anthropic, mistral or openai_compatible
BASEDB_AI_MODEL=mistral-small-latest
MISTRAL_API_KEY=…               # or BASEDB_AI_API_KEY

The key is read from BASEDB_AI_API_KEY, or failing that from the provider’s usual name (OPENAI_API_KEY, ANTHROPIC_API_KEY, MISTRAL_API_KEY).

Azure, a gateway, a local model

BASEDB_AI_PROVIDER=openai_compatible sends the calls, in OpenAI’s format, to the address in BASEDB_AI_BASE_URL: everything that comes before /chat/completions, parameters included. BASEDB_AI_HEADERS adds to every call the headers this server requires, as a JSON object.

# Azure OpenAI: the deployment name as the model, the key in the api-key header
BASEDB_AI_PROVIDER=openai_compatible
BASEDB_AI_MODEL=mon-deploiement
BASEDB_AI_BASE_URL=https://ma-ressource.openai.azure.com/openai/v1
BASEDB_AI_HEADERS='{"api-key":"…"}'

# Azure’s older per-deployment form: the parameter stays after the path
BASEDB_AI_BASE_URL=https://ma-ressource.openai.azure.com/openai/deployments/mon-deploiement?api-version=2024-10-21

# A model served by Ollama, without a key
BASEDB_AI_PROVIDER=openai_compatible
BASEDB_AI_MODEL=llama3.1
BASEDB_AI_BASE_URL=http://ollama:11434/v1

With openai_compatible, the key is optional: if BASEDB_AI_API_KEY is given, it goes out as Authorization: Bearer. A header from BASEDB_AI_HEADERS replaces the key’s — a gateway that wants its own Authorization, for example.

BASEDB_AI_BASE_URL and BASEDB_AI_HEADERS also serve the other three providers, reached through a gateway: for anthropic, the address is everything that comes before /messages. These two variables go with the environment’s provider, and only that one: a workspace that has chosen another one receives neither the address, nor the headers, nor the key. On startup, the API logs the provider it settled on, and flags an invalid address or JSON object.

An internal gateway whose TLS certificate is self-signed, or a company proxy that re-signs the traffic, makes the calls fail: BASEDB_AI_PROVIDER_SSL_VERIFY=false stops verifying the certificate of this provider only — every other outgoing call of the instance, and the provider a workspace may have chosen, stay verified. Startup reports it. Since the key goes out with every call, reserve this for a network you control.

The AI option on a field

AI is not a field type but an option: the AI switch in a field’s form — text, long text, URL, number, single select, boolean, date — has it filled by a model, from a prompt that cites other columns:

Summarize {{Notes}} in one sentence.
Category of {{Description}} among the list’s choices.

• The field is computed as soon as the row exists, then every time a cited column changes — and, if you want, on a schedule (every 15 minutes at most).
• The column keeps its type: an answer where nothing can be read in that type (a number that cannot be found, a choice that does not exist) is rejected rather than written.
• Turning the option off makes the field editable by hand again, values kept.
• The cited values go to the provider: turning it on requires explicit consent.

BASEDB_AI_FIELD_QUOTA caps these computations per hour and per workspace (300 by default).

In an automation

An automation can ask AI in one of its steps: a prompt that cites the row and the previous steps, an answer read in the chosen type, which the following steps write, send or cite. Same rules as for a field: consent when saving, only what the prompt cites goes out, each call logged and counted toward BASEDB_AI_FIELD_QUOTA.

Drafts and Copilot

• Drafts: describe a table or a formula in one sentence, and receive a proposal to review. Only labels, types and the sentence you typed go out — no cell values.
• Templates: describe a whole base — “tracking my clients’ complaints” — and receive tables, sample rows, views, a dashboard and automations, to refine and then create. Only the sentence goes out. See Base templates.
• Copilot: a conversation about the displayed base. You ask for a filter, a query, columns, a table, a test data set; each proposal arrives as a card and is applied in one click, through the same routes as the forms.

By default, only the structure goes to the provider. The “Allow reading data” box lets Copilot read rows for the conversation (50 at most per read) and answer from them — each read is listed under its answer.

Dashboard Copilot

In the Dashboards section, Copilot proposes questions, changes to the dashboard and values for its filters, to apply in one click. Same rules: without consent, only the structure goes out — tables and fields, the base’s dashboards and questions, the definition of the displayed dashboard’s cards (their questions, their texts) —, never the results or the values chosen in the filters. The “Allow reading data” box adds those values and the cards’ results under the displayed filters, 50 rows at most per read, each one listed under the answer.

Automation Copilot

In the Automations section, Copilot proposes a whole automation — the one on screen, modified, or a new one — which it lays on the editor’s flow, without ever saving it: you review it, then save it. Same rules: without consent, only the structure goes out — tables and fields, the base’s automations, the one on screen, its latest runs without any value, people and Slack channels as placeholders —, and the “Allow reading data” box adds rows that were read, 50 at most per read.

BASEDB_AI_QUOTA caps interactive calls per hour and per workspace (120 by default).