Archivist as Proof of Concept

Archivist is our operable proof of concept—demonstrating our ability to integrate several AI technologies into a functional application that solves real business needs.

Built as a working example rather than a fully polished enterprise platform, it showcases our approach to secure, local AI deployment while illustrating some foundational components from which we can develop custom solutions tailored to your specific business needs. The Business Pressure Businesses face mounting pressure to adopt AI technologies to accelerate customer turnaround times, reduce operational costs, and respond to growing demand from customers and stakeholders.

However, this adoption is often hindered by data privacy policies that prohibit uploading sensitive documents to cloud services, restrictive usage limits and rate-limiting on platforms like ChatGPT and Claude, and capabilities that simply don’t match real-world business needs—such as the inability to process large document collections or maintain context across multiple files that are essential for comprehensive business analysis. Enter Archivist

100% Local Operation: Built to run entirely within your business environment, ensuring private documents never leave your secure perimeter
No Usage Limits: Operate without rate-limiting or usage restrictions that cause business delays—process as many documents and queries as needed, when you need them
No Subscription Fees: No cost to use the fully functional base version. A paid license enables users to upload and use more newer and more powerful AI models as they are released.
Advanced Document Management: Purpose-built capabilities that major cloud AI providers lack:
Convert business documents into AI-ready formats
Store documents in a persistent vector database for instant, ongoing retrieval
Advanced metadata operations provide granular control over file organization, enabling users to create project-specific groups, manage case files, organize course materials, and more
Precise citations and traceability to source documents ensure all generated outputs can be validated for accuracy
Data-export capabilities prevent vendor lock-in by preparing data for use with other AI platforms

IBM has emerged as a leader in open-sourcing cutting-edge AI technologies specifically designed for business applications, releasing powerful tools that enterprises can deploy securely within their own infrastructure. IBM has leveraged their unimaginable access to business documents and paperwork to train AI models for various business-related use cases.

At Integral, we recognized the opportunity to demonstrate our expertise by integrating these separate IBM technologies into a cohesive, user-friendly application that unlocks their advanced capabilities and showcases the potential of combining best-in-class open source components into practical business solutions.

Granite We selected the IBM Granite 3.2-2B model as the base model for Archivist for several reasons:

Laptop-Ready Performance: At just 2 billion parameters, Granite 3.2 runs efficiently on standard business laptops without requiring specialized hardware or cloud resources
Enterprise-Grade Responses: Delivers professional, business-appropriate outputs while maintaining the accuracy and detail required for serious business applications
Instruct Architecture: Specifically fine-tuned to follow detailed instructions and respond to complex prompts, making it ideal for understanding and executing the nuanced queries typical in business document retrieval scenarios
Optimized for RAG: Built with advanced prompting capabilities designed specifically for Retrieval Augmented Generation applications, allowing it to seamlessly integrate retrieved document context with generated responses (perfect for Archivist)
Chain-of-Thought Reasoning: Enables the model to break down complex questions step-by-step, providing more accurate and explainable answers when working with business documents
Extensive Context Window: 128K token context length allows processing of large documents or multiple document sections simultaneously, maintaining coherence across lengthy business materials

Our value-add to the IBM Granite model with Archivist is that we created a user-friendly interface for users to toggle between response modes (standard, RAG, chain of thought), response lengths, and RAG response styles – all through the use of simple radio buttons in the UI. This empowers non-technical business users to unlock these advanced features without having to know anything about special prompt tokens.

Also, more than just toggling buttons, we would like to point out that, in RAG mode, the entire pipeline is modified. In RAG mode, the Granite system prompt expects that the retrieved chunks be inserted into the middle of the system prompt and not simply stuffed into the user response. Archivist is developed to operate accordingly and switch pipelines to get the most out of the Granite model – all without any awareness needed by the end user. Docling

Docling is IBM’s text conversion package. It can convert business documents in PDF, DOCX, XLSX, PPTX, CSV, and others to plaintext and markdown which is ideal for input into AI systems.
Docling itself uses multiple AI models to do things like identify page layouts and table structure.
Docling contains multiple settings that can be configured “under the hood” to convert documents using different technologies which yield different results, and processing times, for different types of documents.

Our value-add to the IBM Docling package with Archivist is that we created a user-friendly interface for users to toggle between processing modes through the use of simple radio buttons in the UI. We tested and configured different combinations of settings to give users control over the speed vs. accuracy tradeoff. This empowers non-technical business users to unlock these advanced features with a mouse click. SmolDocling

SmolDocling was co-created by IBM and Hugging Face specifically for document understanding, making it ideal for extracting insights from business documents, PDFs, charts, and diagrams
It can be used to generated descriptions of images within documents, enabling comprehensive content indexing and searchability across both text and visual elements
Can work within the IBM Docling package to yield descriptions of images to go alongside the text converted by Docling
Can operate standalone and provide end-to-end vision-to-text conversion of business documents

Our value-add to the SmolDocling model with Archivist is that we created the pipelines for SmolDocling to annotate images within Docling and for SmolDocling to operate standalone such that users can enable these bleeding-edge technologies with the click of their mouse.

Other Technical Highlights

Optimized for Business Hardware: Utilizes llama.cpp for efficient CPU inference to smoothly on standard business laptops and consumer devices without requiring expensive GPU infrastructure
Multi-Engine Integration: Combines multiple specialized AI engines—llama.cpp for text processing, Hugging Face Transformers for Whisper voice transcription and SmolVLM vision capabilities—demonstrating our ability to orchestrate diverse AI technologies into a unified solution
Commercial-Grade Packaging: Enhanced the core application with professional licensing, security validation, and Microsoft Store distribution, showcasing our capability to transform open-source AI technologies into market-ready commercial products
Flexible Deployment Architecture: Built on Gradio for rapid prototyping, demonstrating our preferred enterprise deployment model—networked AI applications accessible through browsers without individual installations. While Gradio suited this proof of concept, we typically implement purpose-built interfaces and deploy on dedicated hardware for production environments, delivering significant performance improvements and eliminating the overhead limitations of packaged consumer applications.

Take the Next Step

Get the App

Integrate AI into your business