What's new links for OAC and 26ai database

The Oracle Analytics Cloud (OAC) "what's new" section is here: 

https://docs.oracle.com/en/cloud/paas/analytics-cloud/acswn/index.html

It was updated with May release.

The 26ai database "New Features" doc has a May update as well

https://docs.oracle.com/en/database/oracle/oracle-database/26/nfcoa/oracle-ai-database-26ai-new-features-guide.pdf

Installing Oracle Analytics Server (OAS) 2026 on my Windows

This is a practice installation of OAS 2026 on my unsupported Windows 11 Laptop. Usually, I prefer a Linux installation for practical purposes, yet I also do a practice one on my laptop.

If you install on a laptop like me, you might want to check here before installation.

The download page is here, or the edelivery site here. If you use edelivery, don't forget the Java Developers Kit 21. In 21 the JRE is included in JDK. 

After installing JDK, you should have at least the 2 jar files (3 on Linux):

This is the installation guide here.

The basic steps:

• Install JDK 21 if needed
• Install fmw 14.1.2.0.0
• Install Oracle Analytics Server 2026
• Configure Oracle Analytics Server 2026

The short version:

• Download and install JDK from here. I prefer installing in a folder without spaces in the name like c:\JAVA and not the default C:\Program Files\.... that has spaces.
• Install fmw 14.1.2.0.0  
• java.exe -jar fmw_14.1.2.0.0_infrastructure.jar
• Install OAS 2026 in the same Oracle Home 
• java.exe -jar Oracle_Analytics_Server_2026_Windows.jar
• Configure with RCU set up as separate step (as many prefer) or as part of the general Configure flow. (described here)
• RCU: ORACLE_HOME\oracle_common\bin\RCU.bat don't use the default option there (see below)
• Configure: ORACLE_HOME\bi\bin\config.cmd
• Download and configure the Remote Console from this Git

If you install the RCU on the supported Oracle 26ai, you might get "ERROR - RCU-6080 Global prerequisite check failed for the specified database". Check note KB876701 "Creating Repository Schemas for Oracle Analytics Server Errors With 'RCU-6080' " on support site for a patch or workaround.

Extra steps (not covered here):

Enable Users to Export to Excel: https://docs.oracle.com/en/middleware/bi/analytics-server/administer-oas/enable-users-export-excel.html

Enable the Workbook Email Scheduler: https://docs.oracle.com/en/middleware/bi/analytics-server/administer-oas/enable-workbook-email-scheduler.html

Enable Custom Java Script for Actions: https://docs.oracle.com/en/middleware/bi/analytics-server/administer-oas/enable-custom-javascript-actions.html

Configure Advanced System Settings: https://docs.oracle.com/en/middleware/bi/analytics-server/administer-oas/configure-advanced-system-settings.html

The Longer version:

After installing JDK 21, I can install FMW. On Windows run CMD as Administrator.

c:\JAVA\jdk-21.0.10\bin\java.exe -jar fmw_14.1.2.0.0_infrastructure.jar

The installer UI starts

And you should set the Oracle_Home of the installation.

Mine was C:\Oracle\Ora_Home1

Next I install the OAS software 

    c:\JAVA\jdk-21.0.10\bin\java.exe -jar Oracle_Analytics_Server_2026_Windows.jar

in the same home.

You will get Visual C++ installation request

"Setup Failed" usually means this is installed already.

If you want to set RCU as a separate step, please note

The default option of "Oracle Database enabled for edition-based redefinition" doesn't let you select Analytics in the next screen  

The regular Oracle database as Database Type, works fine:

In my case I used the RCU creation as part of the OAS config and it worked fine. 

    ORACLE_HOME\bi\bin\config.cmd

You will see a welcome screen

Next you can select Analytics and Publisher

As Oracle recommends, I changed the Domain Home to a separate location, unlike the default.

Please note you have to create the folder in advance, or you will get the "Domain does not exist"error:

If you used the RCU utility, select to use existing Schemas

If not, create one now:

Next is the port range

And, if you have a previous jar file, you can use it as default.

Run 

In my case this was Successful.

In the past I had few cases when the config timed out on the startup phase. What solved the issue for me then was to startup manually the OAS services without closing the UI window and let the config continue. 

Don't forget to save all the link at the end:

When you try to Open Console, you will discover that you need to download and configure the Remote Console from this Git.

It looks a bit differant.

Let's do the basic task of adding users with the Remote Console.

On the left panel select the last option of Security Tree

 

Select Realms, next press myrealm

Now you can create users

And see them in OAS:

Thanks to Roy Leeram, for the heads-up.

Finding Critical Patch Update (CPU) Advisor For Oracle Analytics Server and Oracle Business Intelligence on oracle support

If you are looking for Critical Patch Update (CPU) Advisor For Oracle Analytics Server and Oracle Business Intelligence (was Document ID 2832967.2) it is now named KA973 and is here now: 

https://support.oracle.com/support/?kmContentId=11137874

It is always a good practice to check here https://www.oracle.com/security-alerts/

Using Hugging Face ONNX models in Oracle 26ai and Fixing the ONNX

 To create Vector and use RAG in Oracle 23/26ai we can encode data using api or load ONNX formatted model and create datamining model in the DB using DBMS_VECTOR.LOAD_ONNX_MODEL (or with OML4Py). 

The ONNX option, usually, has better performance.

*A little shorter syntax might have been using LOAD_ONNX_MODEL_CLOUD. It doesn't influence the core of this post.

I had problems loading few such onnx models from HuggingFace, so I tried to compare one to a default model that is provided by Oracle, all_MiniLM_L12 (see Now Available! Pre-built Embedding Generation model for Oracle Database 26ai). 

I used Oracle Code Assist (with Cline and Codex) to compare Oracle provided version and Hugging Face version and write major portions of this blog.

Intro - How I started:

Uploaded relevant ONNX file to OCI bucket. Created PAR (Pre Authenticated Request) for the bucket (or the file). 

Used DBMS_CLOUD.CREATE_CREDENTIAL to connect it.

BEGIN

  DBMS_CLOUD.CREATE_CREDENTIAL(

    credential_name => 'OCI_OBJ_CRED',

    user_ocid       => 'ocid1.user.oc1… ',

    tenancy_ocid    => 'ocid1.tenancy.oc1… ',

    private_key     => 'THEkeyString==',

    fingerprint     => '57:45:f…'

  );

END;

/

Created a link to the file using DBMS_CLOUD.GET_OBJECT 

BEGIN

  DBMS_CLOUD.GET_OBJECT(

    credential_name => 'OCI_OBJ_CRED',

    object_uri      => 'https://objectstorage.eu-frankfurt-1.oraclecloud.com/p …

                        /all_MiniLM_L12.onnx',

    directory_name  => 'DATA_PUMP_DIR',

    file_name       => 'all_MiniLM_L12.onnx'

  );

END;

/

Loaded the model using DBMS_VECTOR.LOAD_ONNX_MODEL

BEGIN

  DBMS_VECTOR.LOAD_ONNX_MODEL(

    model_name => 'all_MiniLM_L12',

    directory  => 'DATA_PUMP_DIR',

    file_name  => 'all_MiniLM_L12.onnx'

  );

END;

/

It worked well for Oracle’s version of all_MiniLM_L12.onnx but gave me error with the version I downloaded from Hugging Face. 

Got:

ORA-54408: The "input.input_text" field in the JSON metadata is not a model input.

ORA-06512: at "SYS.DBMS_VECTOR", line 3033

ORA-06512: at "SYS.DBMS_DATA_MINING", line 369

ORA-06512: at "SYS.DBMS_DATA_MINING", line 5795

ORA-06512: at "SYS.DBMS_VECTOR", line 3025

ORA-06512: at line 2

https://docs.oracle.com/error-help/db/ora-54408/

Error at Line: 1 Column: 1

After fixing the second onnx version as described below, this version of LOAD_ONNX_MODEL worked:

BEGIN

  DBMS_VECTOR.LOAD_ONNX_MODEL(

    directory  => 'DATA_PUMP_DIR',

    file_name  => 'all_MiniLM_L12_v2.onnx',

    model_name => 'all_MiniLM_L12_v2',

    metadata   => JSON('{

      "function":"embedding",

      "embeddingOutput": "sentence_embedding",

      "input":{"input_text": ["DATA"]}

    }')

  );

END;

/

Core solution:  ONNX + Oracle 23/26ai Embeddings: A Practical Field Guide

If you load custom ONNX embedding models with `DBMS_VECTOR.LOAD_ONNX_MODEL`, model validity alone is **not** enough.

You must align three things:

1. ONNX tensor contract (input/output names, dtypes, ranks)

2. Oracle metadata JSON schema (`embeddingOutput`, `input` mapping)

3. SQL alias mapping used in `VECTOR_EMBEDDING ... USING ... AS ...`

This guide summarizes a real troubleshooting path and the final repeatable pattern.

## 1) Why this path matters

In Oracle 23/26ai, you can generate embeddings either via:

• external APIs
• using OML4Py 
• by loading ONNX models directly into DB (this post)

The ONNX route often has better control and performance but is strict about model contract compatibility.

## 2) Baseline sanity check (known-good sample)

Use Oracle’s provided MiniLM sample as your baseline:

- `all_MiniLM_L12_v2_GOOD.onnx`

If baseline works and your custom model fails, your environment is likely fine, and the issue is model contract/metadata.

## 3) Error progression and what it usually means

### ORA-54446

`The embedding input tensor "input_ids" has an invalid element data type.`

**Meaning:** token-ID input model (`input_ids`, `attention_mask`, etc.) is being loaded into a text-input embedding path.

### ORA-54408

`The "input.<name>" field in the JSON metadata is not a model input.`

**Meaning:** metadata input mapping key doesn’t match real ONNX input tensor name.

### ORA-54435

`Invalid JSON field: "embedding"`

**Meaning:** wrong metadata key for your DB release. Use `embeddingOutput`.

### ORA-54442

`Incorrect batch size location in tensor "input_text"`

**Meaning:** input rank/shape is wrong for Oracle batch expectation.

Common failing shape: `input_text: STRING [1, num_sentences]`.

Common working shape: `input_text: STRING ['batch_size']`.

## 4) Working ONNX contract (MiniLM sample)

- input name: `input`

- input type/shape: `STRING ['batch_size']`

- output name: `embedding`

- output type/shape: `FLOAT ['batch_size', 384]`

## 5) Canonical metadata JSON (for this DB style)

```json

{

  "function": "embedding",

  "embeddingOutput": "embedding",

  "input": {"input": ["DATA"]}

}

```

Notes:

- `embeddingOutput` must match a real ONNX output tensor name.

- `input` mapping key must match a real ONNX input tensor name.

- `DATA` is the SQL alias used in scoring (`USING ... AS DATA`).

## 6) General ORA-54442 remediation (Good -> Good2 patch)

If your converted model fails with ORA-54442 and has rank-2 text input, patch input to rank-1 batch.

```python

import onnx

src = r"C:\path\to\your_model_Good.onnx"

dst = r"C:\path\to\your_model_Good2.onnx"

m = onnx.load(src)

i = m.graph.input[0]

# Force rank-1 batch input for text tensor

del i.type.tensor_type.shape.dim[:]

d = i.type.tensor_type.shape.dim.add()

d.dim_param = "batch_size"

onnx.checker.check_model(m)

onnx.save(m, dst)

print("saved", dst)

```

Quick validation:

```python

import onnx

m = onnx.load(r"C:\path\to\your_model_Good2.onnx")

i = m.graph.input[0]

o = m.graph.output[0]

print("IN", i.name, [d.dim_param if d.dim_param else d.dim_value for d in i.type.tensor_type.shape.dim])

print("OUT", o.name, [d.dim_param if d.dim_param else d.dim_value for d in o.type.tensor_type.shape.dim])

```

## 7) ONNX pre-check script (before upload)

```python

# save as: check_oracle_adb_onnx.py

from pathlib import Path

import sys

import onnx

TYPES = {

    1: "FLOAT", 2: "UINT8", 3: "INT8", 4: "UINT16", 5: "INT16",

    6: "INT32", 7: "INT64", 8: "STRING", 9: "BOOL", 10: "FLOAT16",

    11: "DOUBLE", 12: "UINT32", 13: "UINT64", 16: "BFLOAT16",

}

def shape_list(value_info):

    t = value_info.type.tensor_type

    return [d.dim_param if d.dim_param else d.dim_value for d in t.shape.dim]

def dtype_name(value_info):

    return TYPES.get(value_info.type.tensor_type.elem_type, str(value_info.type.tensor_type.elem_type))

def check(path: Path) -> int:

    m = onnx.load(str(path))

    print(f"\n=== {path.name} ===")

    print("Inputs:")

    for i in m.graph.input:

        print(f"  - {i.name}: {dtype_name(i)}, shape={shape_list(i)}")

    print("Outputs:")

    for o in m.graph.output:

        print(f"  - {o.name}: {dtype_name(o)}, shape={shape_list(o)}")

    errors = []

    if len(m.graph.input) != 1:

        errors.append("Expected exactly 1 model input.")

    else:

        inp = m.graph.input[0]

        if dtype_name(inp) != "STRING":

            errors.append(f"Input dtype should be STRING, got {dtype_name(inp)}.")

        if len(shape_list(inp)) != 1:

            errors.append(f"Input rank should be 1 [batch_size], got rank {len(shape_list(inp))}.")

    if len(m.graph.output) < 1:

        errors.append("Expected at least 1 model output.")

    else:

        out = m.graph.output[0]

        if dtype_name(out) not in ("FLOAT", "FLOAT16", "BFLOAT16"):

            errors.append(f"Output dtype should be float-like, got {dtype_name(out)}.")

        if len(shape_list(out)) != 2:

            errors.append(f"Output rank should be 2 [batch_size, dim], got rank {len(shape_list(out))}.")

    if errors:

        print("\nRESULT: FAIL")

        for e in errors:

            print("  *", e)

        return 1

    print("\nRESULT: PASS (heuristically compatible)")

    return 0

if __name__ == "__main__":

    if len(sys.argv) != 2:

        print("Usage: py -3 check_oracle_adb_onnx.py <model.onnx>")

        sys.exit(2)

    sys.exit(check(Path(sys.argv[1])))

```

Run:

```bash

py -3 check_oracle_adb_onnx.py "C:\path\to\your_model.onnx"

```

## 8) Working SQL pattern (MiniLM sample)

```sql

BEGIN

  DBMS_CLOUD.GET_OBJECT(

    credential_name => 'OCI_OBJ_CRED',

    object_uri      => 'https://objectstorage.<region>.oraclecloud.com/n/<ns>/b/<bucket>/o/all_MiniLM_L12_v2_GOOD.onnx',

    directory_name  => 'DATA_PUMP_DIR',

    file_name       => 'all_MiniLM_L12_v2_GOOD.onnx'

  );

END;

/

```

```sql

BEGIN

  DBMS_VECTOR.LOAD_ONNX_MODEL(

    directory  => 'DATA_PUMP_DIR',

    file_name  => 'all_MiniLM_L12_v2_GOOD.onnx',

    model_name => 'ALL_MINILM_L12_V2_GOOD',

    metadata   => JSON('{

      "function": "embedding",

      "embeddingOutput": "embedding",

      "input": {"input": ["DATA"]}

    }')

  );

END;

/

```

```sql

SELECT VECTOR_EMBEDDING(ALL_MINILM_L12_V2_GOOD USING 'Hello World' AS DATA) AS emb

FROM dual;

```

## 9) Operational checklist

1. Inspect ONNX input/output names before load.

2. Prefer explicit metadata over defaults.

3. Ensure `embeddingOutput` key matches your DB release.

4. If ORA-54442 appears, patch rank-2 text input to rank-1 batch.

5. Treat ORA errors as contract diagnostics, not random failures.

## Final takeaway

Reliable in-DB embeddings depend on strict interface alignment:

**ONNX tensor contract + Oracle metadata contract + SQL alias mapping**

Once aligned, `LOAD_ONNX_MODEL` and `VECTOR_EMBEDDING` become repeatable and predictable.

Alternative approach of  OML4Py:  Running Hugging Face models inside Oracle 23ai with ONNX and OML4Py by fsarcosdb