Merge branch 'main' into pob-app-integration-and-automation

Author: MarcelStraka

ai/generative-ai-service/smart-invoice-extraction/README.md

@@ -0,0 +1,107 @@
+# 🧾 Invoice Analysis Plus
+
+An intelligent invoice data extractor built with **OCI Generative AI**, **LangChain**, and **Streamlit**. Upload any invoice PDF and this app will extract structured data like REF. NO., POLICY NO., DATES, etc. using multimodal LLMs.
+
+---
+
+## 🚀 Features
+
+- 🔍 Automatically identifies key invoice headers using OCI Vision LLM (LLaMA 3.2 90B Vision)
+- 🤖 Lets you choose what elements to extract (with type selection)
+- 🧠 Leverages a text-based LLM (Cohere Command R+) for context-aware value extraction
+- 🧪 Outputs data in clean **JSON** and saves to **CSV**
+- 🖼️ Uses image-based prompt injection for high accuracy
+
+---
+
+## 🛠️ Tech Stack
+
+| Tool                | Usage                                   |
+|---------------------|------------------------------------------|
+| 🧠 OCI Generative AI | Vision + Text LLMs for extraction       |
+| 🧱 LangChain         | Prompt orchestration and LLM chaining   |
+| 📦 Streamlit         | Interactive UI and file handling        |
+| 🖼️ pdf2image         | Convert PDFs into JPEGs                 |
+| 🧾 Pandas            | CSV creation & table rendering          |
+| 🔐 Base64            | Encodes image bytes for prompt injection|
+
+---
+
+## 🧠 How It Works
+
+1. **User Uploads Invoice PDF**  
+   The file is uploaded and converted into an image using `pdf2image` (Ensure you upload one page documents ONLY)
+
+2. **Initial Header Detection (LLaMA-3.2 Vision)**  
+   The first page is passed to the multimodal LLM which returns a list of fields that are likely to be useful (e.g., "Policy No.", "Amount", "Underwriter").
+
+3. **User Selects Fields and Types**  
+   A UI allows the user to pick 3 fields from the detected list, and specify their data types (Text, Number, etc.).
+
+4. **Prompt Generation (Cohere Command R+)**  
+   The second LLM generates a custom system prompt to extract those fields as JSON.
+
+5. **Full Invoice Extraction (LLaMA-3.2 Vision)**  
+   Each page image is passed into the multimodal LLM using the custom prompt, returning JSON values for the requested fields.
+
+6. **Data Saving & Display**  
+   All data is shown in a `st.dataframe()` and saved to CSV.
+
+---
+
+## 📁 File Structure
+
+```bash
+.
+├── app.py               # Main Streamlit app
+├── requirements.txt     # Python dependencies
+└── README.md            # This file
+```
+
+---
+
+## 🔧 Setup
+
+1. **Clone the repository**
+
+```bash
+git clone <repository-url>
+cd <repository-folder>
+```
+
+2. **Install dependencies**
+
+```bash
+pip install -r requirements.txt
+```
+
+3. **Run the app**
+
+```bash
+streamlit run app.py
+```
+
+> ⚠️ **Important Configuration:**
+>
+> - Replace all instances of `<YOUR_COMPARTMENT_OCID_HERE>` with your actual **OCI Compartment OCID**
+> - Ensure you have access to **OCI Generative AI Services** with correct permissions
+> - Update model IDs in the code if needed:  
+>   - Vision model: `meta.llama-3.2-90b-vision-instruct`  
+>   - Text model: `cohere.command-r-plus-08-2024`
+
+---
+
+## 📁 Output Sample
+
+```json
+[
+  {
+    "REF. NO.": "IN123456",
+    "INSURED": "Acme Corp",
+    "POLICY NO.": "POL987654",
+    "File Name": "invoice1.pdf",
+    "Page Number": 1
+  },
+  ...
+]
+```

ai/generative-ai-service/smart-invoice-extraction/SmartInvoiceExtraction.py

@@ -0,0 +1,185 @@
+import pandas as pd
+import json
+from langchain.chains.llm import LLMChain
+from langchain_core.prompts import PromptTemplate
+import streamlit as st
+from langchain_community.chat_models.oci_generative_ai import ChatOCIGenAI
+from langchain_core.messages import HumanMessage, SystemMessage
+import base64
+from pdf2image import convert_from_bytes
+import io
+
+# Helper function to convert a list of images into byte arrays for further processing
+def save_images(images, output_format="JPEG"):
+    image_list = []
+    for image in images:
+        img_byte_arr = io.BytesIO()
+        image.save(img_byte_arr, format=output_format)
+        img_byte_arr.seek(0)
+        image_list.append(img_byte_arr)
+    return image_list
+
+# Helper function to encode an image to base64 for sending to LLM
+def encode_image(image_path):
+    with open(image_path, "rb") as image_file:
+        return base64.b64encode(image_file.read()).decode("utf-8")
+
+# Save extracted data to a CSV file and show success message in Streamlit
+def save_to_csv(data, file_name="extracted_data.csv"):
+    df = pd.DataFrame(data)
+    df.to_csv(file_name, index=False)
+    st.success(f"Data saved to {file_name}")
+
+# Extract key headers from the first image of a PDF invoice
+def extractor(image_list):
+    # Replace this with your own compartment ID
+    compID = "<YOUR_COMPARTMENT_OCID_HERE>"
+
+    # Load a multimodal LLM for invoice header analysis
+    llm = ChatOCIGenAI(
+        model_id="meta.llama-3.2-90b-vision-instruct",  # Replace with your model ID
+        compartment_id=compID,
+        model_kwargs={"max_tokens": 2000, "temperature": 0}
+    )
+
+    # Encode the first page as base64
+    encoded_frame = base64.b64encode(image_list[0].getvalue()).decode("utf-8")
+
+    with st.spinner("Extracting the key elements"):
+        # Provide system instruction to extract headers from invoice
+        system_message = SystemMessage(
+            content="""Given this invoice, extract in list format, all the headers that can be needed for analysis
+            For example: [\"REF. NO.\", \"INSURED\", \"REINSURED\", \"POLICY NO.\", \"TYPE\", \"UNDERWRITER REF. NO.\", \"PERIOD\", \"PARTICULARS\", \"PPW DUE DATE\"]
+            Return the answer in a list format, and include nothing else at all in the response.
+            """
+        )
+
+        # Human message includes the image
+        human_message = HumanMessage(
+            content=[
+                {"type": "text", "text": "This is my invoice"},
+                {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encoded_frame}"}},
+            ]
+        )
+
+        # Invoke the LLM and extract elements
+        ai_response = llm.invoke(input=[human_message, system_message])
+        st.caption("Here are some key elements you may want to extract")
+        return eval(ai_response.content)
+
+# Main Streamlit app function
+def invoiceAnalysisPlus():
+    st.title("Invoice Data Extraction")
+
+    with st.sidebar:
+        st.title("Parameters")
+        # Replace with your own compartment ID
+        compID = "<YOUR_COMPARTMENT_OCID_HERE>"
+        user_prompt = st.text_input("Input the elements you are looking to extract here")
+        st.caption("Our AI assistant has extracted the following key elements from the invoice. Please select the elements you wish to extract.")
+
+    uploaded_file = st.file_uploader("Upload your invoices here:", type=["pdf"])
+
+    if uploaded_file is not None:
+        with st.spinner("Processing..."):
+            # Convert PDF to image list
+            if uploaded_file.type == "application/pdf":
+                images = convert_from_bytes(uploaded_file.read(), fmt="jpeg")
+            else:
+                images = [convert_from_bytes(uploaded_file.read(), fmt="jpeg")[0]]
+
+            # Save as byte streams
+            image_list = save_images(images)
+
+        # Load both image-based and text-based LLMs
+        llm = ChatOCIGenAI(
+            model_id="meta.llama-3.2-90b-vision-instruct",  # Replace with your model ID
+            compartment_id=compID,
+            model_kwargs={"max_tokens": 2000, "temperature": 0}
+        )
+        llm_for_prompts = ChatOCIGenAI(
+            model_id="cohere.command-r-plus-08-2024",  # Replace with your model ID
+            compartment_id=compID,
+            model_kwargs={"max_tokens": 2000, "temperature": 0}
+        )
+
+        # Select box UI for user to pick elements and their data types
+        data_types = ["Text", "Number", "Percentage", "Date"]
+        elements = []
+
+        if "availables" not in st.session_state:
+            st.session_state.availables = extractor(image_list)
+
+        for i in range(3):  # Max 3 fields
+            col1, col2 = st.columns([2, 1])
+            with col1:
+                name = st.selectbox(f"Select an element {i+1}", st.session_state.availables, key=f"name_{i}", index=i)
+            with col2:
+                data_type = st.selectbox(f"Type {i+1}", data_types, key=f"type_{i}")
+                elements.append((name, data_type))
+
+        # Generate appropriate prompt based on selected or input fields
+        if elements:
+            system_message_cohere = SystemMessage(
+                content=f"""
+                Based on the following set of elements {elements}, with their respective types, extract their values and respond only in valid JSON format (no explanation):
+                {', '.join([f'- {e[0]}' for e in elements])}
+                For example:
+                {{
+                    {elements[0][0]}: "296969",
+                    {elements[1][0]}: "296969",
+                    {elements[2][0]}: "296969"
+                }}
+                """
+            )
+            ai_response_cohere = system_message_cohere
+        else:
+            system_message_cohere = SystemMessage(
+                content=f"""
+                Generate a system prompt to extract fields based on user-defined elements: {user_prompt}.
+                Output should be JSON only. No other text.
+                """
+            )
+            ai_response_cohere = llm_for_prompts.invoke(input=[system_message_cohere])
+
+        # Extracted data list
+        extracted_data = []
+
+        with st.spinner("Analyzing invoice..."):
+            for idx, img_byte_arr in enumerate(image_list):
+                try:
+                    encoded_frame = base64.b64encode(img_byte_arr.getvalue()).decode("utf-8")
+
+                    if elements:
+                        system_message = ai_response_cohere
+                    else:
+                        system_message = SystemMessage(content=ai_response_cohere.content)
+
+                    human_message = HumanMessage(
+                        content=[
+                            {"type": "text", "text": "This is my invoice"},
+                            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encoded_frame}"}},
+                        ]
+                    )
+
+                    ai_response = llm.invoke(input=[human_message, system_message])
+                    json_start = ai_response.content.find('{')
+                    json_end = ai_response.content.find('}', json_start)
+                    json_data = ai_response.content[json_start:json_end + 1]
+
+                    response_dict = json.loads(json_data)
+                    response_dict["File Name"] = uploaded_file.name
+                    response_dict["Page Number"] = idx + 1
+                    extracted_data.append(response_dict)
+
+                except Exception as e:
+                    st.error(f"Error processing page {idx+1}: {str(e)}")
+
+        # Display and save results
+        if extracted_data:
+            save_to_csv(extracted_data)
+            st.dataframe(pd.DataFrame(extracted_data))
+
+# Run the app
+if __name__ == "__main__":
+    invoiceAnalysisPlus()

ai/generative-ai-service/smart-invoice-extraction/requirements.txt

@@ -0,0 +1,8 @@
+pandas==1.5.3
+json==2.0.9
+langchain==0.0.209
+langchain_community==0.0.1
+streamlit==1.24.0
+oci==2.97.0
+pdf2image==1.16.3
+Pillow==8.4.0

cloud-infrastructure/private-cloud-and-edge/compute-cloud-at-customer/local-llm/README.md

@@ -1,4 +1,4 @@
-*Last Update: 27 November 2024*
+*Last Update: 04.04.2025*
 
 <br><h1 align="center">Local LLM Inferencing and Interaction<br>Using the Ollama Open Source Tool</h1>
 <p align="center"><img align="centre" src="./images/ollama-logo.png" width="10%" style="float:right"/></p>
@@ -21,12 +21,12 @@
 
 By running models locally, you maintain full data ownership and avoid the potential security risks associated with cloud storage. Offline AI tools like Ollama also help reduce latency and reliance on external facilities, making them faster and more reliable.
 
-This article is intended to demonstrate and provide directions to install and create an Ollama LLM processing facility. Despite the fact that Ollama can be run on both personal servers and laptops, this installation is aimed at the Oracle Compute Cloud@Customer (C3) and Private Cloud Appliance (PCA) to capitalize on more readily available resources to increase performance and processing efficiency, especially if large models are used.
+This article is intended to demonstrate and provide directions to install and create an Ollama LLM processing facility. Despite the fact that Ollama can be run on both personal servers and laptops, this installation is aimed at the Oracle Compute Cloud@Customer (C3) to capitalize on more readily available resources to increase performance and processing efficiency, especially if large models are used.
 
 Considerations:
-* A firm grasp of C3/PCA/OCI concepts and administration is assumed.
+* A firm grasp of C3 and OCI concepts and administration is assumed.
 * The creation and integration of a development environment is outside of the scope of this document.
-* Oracle Linux 8 and macOS Sonoma 14.7.1 clients were used for testing but Windows is however widely supported.
+* Oracle Linux 8, macOS Sonoma 14.7.1 and macOS Sequoia 15.3.2 clients were used for testing but Windows is however widely supported.
 
 [Back to top](#toc)<br>
 <br>
@@ -40,17 +40,16 @@ Considerations:
 |----------|----------|
 | Operating system | Oracle Linux 8 or later<br>Ubuntu 22.04 or later<br>Windows<br> |
 | RAM | 16 GB for running models up to 7B. "The rule of thumb" is to have at least 2x memory for the size of the LLM, also allowing for LLMs that will be loaded in memory simultaneously. |
-| Disk space | 12 GB for installing Ollama and basic models. Additional space is required for storing model data depending on the used models. The LLM sizes can be obtained from the "trained models" link in the References section. For example the Llama 3.1 LLM with 405Bn parameters occupy 229GB of disk space |
+| Disk space | 12 GB for installing Ollama and basic models. Additional space is required for storing model data depending on the used models. The LLM sizes can be obtained from the "trained models" link in the References section. For example the Llama 3.1 LLM with 405Bn parameters occupies 229GB of disk space |
 | Processor | Recommended to use a modern CPU with at least 4 cores. For running models of approximately 15B, 8 cores (OCPUs) is recommended. Allocate accordingly |
-| Graphics Processing Unit<br>(optional) | A GPU is not required for running Ollama, but can improve performance, especially when working with large models. If you have a GPU, you can use it to accelerate training of custom models. |
+| Graphics Processing Unit<br>(optional) | A GPU is not required for running Ollama, but can improve performance, especially when working with large models. If you have a GPU, you can use it to accelerate inferencing, training, fine-tuning and RAG (Retrieval Augmented Generation). |
 
 >[!NOTE]
->The GPU options in the Compute Cloud@Customer will be available soon.
+>The C3 now has an NVIDIA L40S GPU expansion option available. Using a 4-GPU VM it is expected that performance acceleration will dramatically improve for LLMs of up to approximately 70bn parameters.
 
 ### Create a Virtual Machine Instance
 
-[C3: Creating an Instance](https://docs.oracle.com/en-us/iaas/compute-cloud-at-customer/topics/compute/creating-an-instance.htm#creating-an-instance)<br>
-[PCA 3.0: Working with Instances](https://docs.oracle.com/en/engineered-systems/private-cloud-appliance/3.0-latest/user/user-usr-instance-lifecycle.html)
+[Creating an Instance](https://docs.oracle.com/en-us/iaas/compute-cloud-at-customer/topics/compute/creating-an-instance.htm#creating-an-instance)<br>
 
 Create a VM in a public subnet following these guidelines:
 
@@ -73,8 +72,7 @@ sudo dnf update
 
 ### Create a Block Storage Device for LLMs
 
-[C3: Creating and Attaching Block Volumes](https://docs.oracle.com/en-us/iaas/compute-cloud-at-customer/topics/block/creating-and-attaching-block-volumes.htm)<br>
-[PCA 3.0: Creating and Attaching Block Volumes](https://docs.oracle.com/en/engineered-systems/private-cloud-appliance/3.0-latest/user/user-usr-blk-volume-create-attach.html)
+[Creating and Attaching Block Volumes](https://docs.oracle.com/en-us/iaas/compute-cloud-at-customer/topics/block/creating-and-attaching-block-volumes.htm)<br>
 
 1.  Create and attach a block volume to the VM
 2.  Volume name `llm-repo`
@@ -106,7 +104,7 @@ export no_proxy
 ```
 
 >[!TIP]
->The `no_proxy` environment variable can be expanded to include your internal domains. It is not required to list IP addresses in internal subnets of the C3/PCA.
+>The `no_proxy` environment variable can be expanded to include your internal domains. It is not required to list IP addresses in internal subnets of the C3.
 
 Edit the `/etc/yum.conf` file to include the following line:
 ```
@@ -181,7 +179,7 @@ The installation comprises the following components:
 <sup><sub>2</sup></sub> See [Ollama documentation](https://github.com/ollama/ollama/tree/main/docs)
 
 >[!IMPORTANT]
->When GPU's become available the NVIDIA and CUDA drivers should be installed. This configuration will also be tested on the Roving Edge Device GPU model.
+>When GPU's are available for use on the C3 the NVIDIA and CUDA drivers should be installed. This configuration will also be tested on the Roving Edge Device GPU model.
 
 ### Installation
 
@@ -324,3 +322,4 @@ Copyright (c) 2025 Oracle and/or its affiliates.
 Licensed under the Universal Permissive License (UPL), Version 1.0.
 
 See [LICENSE](LICENSE) for more details.
+