Run a local LLM (Ollama) inside Oracle APEX

For sensitive workloads you may not want to send data to a cloud LLM. Ollama runs open-weight models on your own machine or server and exposes them over a local HTTP API. Oracle APEX can call that API from PL/SQL — or, from APEX 26.1, use it as a built-in AI provider — so inference stays on infrastructure you control.

Run Ollama

Install Ollama and pull a small instruct model (llama3.2 is the 3B model, about 2 GB). The Windows app runs in the background after install; on Linux the server usually runs as a systemd service — if it isn't running, start it with ollama serve. Either way it serves an HTTP API on http://localhost:11434.

# Install Ollama from https://ollama.com/download (it runs in the background), then:
ollama pull llama3.2
ollama run llama3.2 "Say hello in one sentence"
curl.exe http://localhost:11434/api/version

Allow the database to make the call

Outbound network calls are blocked until a network ACL allows them. APEX_WEB_SERVICE and the APEX AI features make the call as the APEX engine schema (for example APEX_260100 on APEX 26.1) — not as APEX_PUBLIC_USER and not as your parsing schema. Many installs already grant it during setup, so check first. Connect as a DBA to the PDB where APEX is installed — Oracle AI Database 26ai is always a container database, and CDB$ROOT has no APEX, so the grant below fails there with PLS-00201 on APEX_APPLICATION. Replace FREEPDB1 with your PDB name:

ALTER SESSION SET CONTAINER = FREEPDB1;

SELECT host, lower_port, upper_port, principal, privilege
FROM   dba_host_aces
ORDER  BY host, principal;

If there is no matching entry, grant connect on just the Ollama host and port. Run this as SYS (or another user allowed to run DBMS_NETWORK_ACL_ADMIN) in the same APEX PDB session; APEX_APPLICATION.g_flow_schema_owner resolves to the current APEX engine schema, so the script survives APEX upgrades:

BEGIN
  DBMS_NETWORK_ACL_ADMIN.APPEND_HOST_ACE(
    host       => 'host.docker.internal',
    lower_port => 11434,
    upper_port => 11434,
    ace        => xs$ace_type(privilege_list => xs$name_list('connect'),
                              principal_name => APEX_APPLICATION.g_flow_schema_owner,
                              principal_type => xs_acl.ptype_db));
END;
/

Call the model from PL/SQL

Post to Ollama's /api/generate endpoint. Set "stream": false — the API streams by default — and read the generated text from the response field. Building the body with JSON_OBJECT_T escapes quotes in user input for you.

DECLARE
  l_req    json_object_t := json_object_t();
  l_resp   CLOB;
  l_answer CLOB;
BEGIN
  l_req.put('model',  'llama3.2');
  l_req.put('prompt', 'Summarize in one sentence: order 1001 shipped two days late because of a customs hold.');
  l_req.put('stream', false);

  apex_web_service.set_request_headers(
    p_name_01  => 'Content-Type',
    p_value_01 => 'application/json');

  l_resp := apex_web_service.make_rest_request(
    p_url              => 'http://host.docker.internal:11434/api/generate',
    p_http_method      => 'POST',
    p_body             => l_req.to_clob,
    p_transfer_timeout => 120);

  IF apex_web_service.g_status_code <> 200 THEN
    raise_application_error(-20001,
      'Ollama returned HTTP ' || apex_web_service.g_status_code || ': ' ||
      dbms_lob.substr(l_resp, 500, 1));
  END IF;

  -- The generated text is in the "response" field
  l_answer := json_object_t.parse(l_resp).get_clob('response');
  dbms_output.put_line(l_answer);
END;
/

For multi-turn conversations use /api/chat with a messages array instead. Ollama also offers an OpenAI-compatible API under http://localhost:11434/v1 (for example /v1/chat/completions), which is handy if your code already speaks the OpenAI format.

Or: register Ollama as an APEX Generative AI Service (APEX 26.1+)

From APEX 26.1, Ollama and OpenAI-compatible endpoints are built-in AI providers. In App Builder go to Workspace Utilities → Generative AI, click Create, choose the Ollama provider, enter the Ollama server URL as the Base URL, set AI Model to llama3.2, give it a Static ID such as LOCAL_OLLAMA, and click Test Connection. The same network ACL applies. Which Ollama API APEX calls depends on the release: APEX 26.1 uses Ollama's native /api/chat and /api/embed; in APEX 26.2 new Ollama services default to the Responses API (Chat Completions is deprecated), selectable under Advanced → Provider API. Ollama added its OpenAI-compatible /v1/responses endpoint in v0.13.3, so upgrade older Ollama installs before using that default. The service can then back the declarative AI features or be called from PL/SQL running in that workspace's apps with APEX_AI:

DECLARE
  l_answer CLOB;
BEGIN
  l_answer := apex_ai.generate(
    p_prompt            => 'Summarize in one sentence: order 1001 shipped two days late because of a customs hold.',
    p_system_prompt     => 'You are a concise assistant for an order-management app.',
    p_service_static_id => 'LOCAL_OLLAMA');
  dbms_output.put_line(l_answer);
END;
/

Check your understanding