The Virtual Intel Technical Assistant (VTA) is a generative AI–powered chatbot that helps Intel technical sellers, partners, and internal users quickly find accurate, source-backed information about Intel products and solutions. Intel built VTA to address a practical challenge: technical sellers were spending too much time searching through thousands of long, complex documents spread across many repositories—manuals, portals, emails, and more. This slowed down their ability to respond to customer questions and made it harder to provide precise, up-to-date answers. VTA uses a retrieval-augmented generation (RAG) approach on top of large language models (initially GPT-4 via the OpenAI API) to: - Accept natural-language questions about Intel products and technologies - Retrieve relevant content from Intel’s internal knowledge bases - Generate concise, grounded answers - Provide direct links and exact content snippets from the original documents The assistant was first rolled out to Intel’s Data Center and AI (DCAI) business unit and is being expanded to other products and business units. Within about six months, usage grew from a few hundred early users to roughly 5,000 users in Intel’s Sales and Marketing Group (SMG) and business units, with a target of scaling to around 20,000 users. In short, VTA is designed to help Intel sellers respond faster and more confidently to customer questions, improve engagement, and support business growth by making Intel’s technical knowledge easier to access and use.

VTA combines retrieval-augmented generation (RAG) with large language models (LLMs) to give Intel sellers grounded, explainable answers instead of generic AI responses. Here’s how the workflow operates: 1. **User asks a question** A seller types a natural-language question into the VTA interface. They don’t need to know where the relevant documents are stored. 2. **Relevant information is retrieved** The RAG system queries Intel’s internal knowledge base, which is populated from diverse data sources such as documents, code, and databases. An offline pipeline pre-processes this content using intelligent parsing and chunking to support both: - **Sparse (text-based) search** via Elasticsearch - **Dense (semantic) search** via Milvus for vector embeddings MongoDB stores structured and unstructured documents in a JSON-like format. All three databases are built with scale-out architectures to handle large and growing data volumes. 3. **Context is built for the LLM** The retrieved snippets are packaged as context and combined with the user’s question into a prompt for the LLM. VTA currently uses GPT-4 via the OpenAI API for generation, but the platform is designed to support multiple LLMs, including internal, fine-tuned models. 4. **The LLM generates a grounded answer** The model produces an answer that is explicitly grounded in the retrieved content. VTA then: - Shows the answer in a simple chat-style UI - Lists the exact content sections used - Provides direct links to the original full documents 5. **Access control and classifications** Content links are tagged with classifications such as Public, NDA-required, Restricted, and Restricted-Secret. This helps sellers share only what is appropriate for each customer relationship and agreement. 6. **Continuous improvement via feedback** Users can rate answers with a star-based scoring system. This feedback is used to: - Identify high-quality answers as a “gold standard” for future fine-tuning - Expose content gaps where documentation is missing or unclear This RAG-based design delivers several key benefits: - **Better factual accuracy and recency** by pulling from Intel’s latest internal repositories instead of relying only on a 6–12 month–old LLM training snapshot - **Reduced hallucinations** because the model is constrained by retrieved, domain-specific context - **Efficient customization** for new domains and use cases without expensive full-model retraining Under the hood, VTA runs on Intel hardware (Intel Xeon processors with Intel Accelerator Engines) and Kubernetes for portability and scalability. Calls to external LLMs like GPT-4 are optional and can be replaced with on-premises, fine-tuned open-source models (e.g., Llama, StarCoder, Mixtral) running on Intel Gaudi-based systems in future versions.

VTA has already shown tangible business impact for Intel’s technical sellers and their customers, even in its initial rollout phase. Key outcomes include: 1. **High and growing adoption** - Started with a few hundred early users - Grew to about **5,000 users** in the Sales and Marketing Group (SMG) and business units within roughly six months - Designed to scale to around **20,000 users** over time This adoption indicates that VTA has become a go-to resource for technical queries inside Intel. 2. **Instant answers instead of hours or days** Before VTA, sellers often spent hours or even days searching across multiple repositories and long documents to answer complex questions. VTA now delivers: - **Instant responses** to many technical questions - Summarized, source-backed answers that sellers can quickly validate and share 3. **Faster design-in and shorter sales cycles** By speeding up access to accurate technical details during product selection and integration, VTA helps: - Accelerate design-in decisions - Support smoother integration discussions - Shorten parts of the sales cycle where technical validation is critical 4. **Improved customer engagement and satisfaction** VTA supports better customer conversations by: - Providing timely, accurate, and attributable information - Giving sellers confidence to address deeper, more complex technical questions - Enabling quick access to original documents when customers need more detail 5. **Higher seller productivity and knowledge democratization** Sellers can now: - Spend less time hunting for information - Handle more complicated queries themselves - Tap into knowledge that was previously siloed in specific teams or experts This democratizes access to Intel’s technical expertise and makes it easier for more sellers to operate at a higher technical level. 6. **Better content and product insights** VTA also feeds insights back into Intel’s content and product strategies by: - Highlighting unanswered or low-quality responses, which reveal documentation gaps - Providing visibility into common design challenges and recurring questions for specific products Looking ahead, Intel plans to extend VTA’s impact by integrating it with additional data sources such as ticketing systems, CRM platforms, and product roadmaps. This will help VTA provide richer context about each customer’s environment, anticipate issues earlier, and fit more seamlessly into everyday sales workflows. Overall, VTA is helping Intel reimagine how technical knowledge is accessed and used in sales, leading to faster responses, more effective customer engagements, and better-informed internal decision-making.