Interview
Vector Databases and the Data Structure of AI ft. MongoDB’s Sahir Azam
- Quality Engineering in Probabilistic Software
- In a probabilistic world, software quality is defined by the "last mile" required to achieve 99.99x reliability.
- Traditional deterministic quality engineering models do not fully apply to AI applications without adaptation.
- High-quality retrieval in enterprise environments depends on embedding models, RAG architecture construction, and merging data with real-time transactional views.
- Mission-critical use cases in conservative enterprises require this high-quality probabilistic output to be adopted.
- Market Shifts and Software Evolution
- Generative AI is enabling "services as software," addressing use cases previously impossible for deterministic systems.
- Software creation is becoming significantly easier, driving an increase in total software volume and data persistence needs.
- Human-computer interaction is shifting toward "ambient agents" that react to signals without intentional human action, impacting robotics and productivity.
- More software generation implies a tailwind for data infrastructure, as state management becomes more complex in agent-driven workflows.
- Case Study: Automotive Industry
- A European automaker implemented audio embedding models to semantically match car noises against known problem corpora.
- The solution reduced complex car diagnosis time from hours to seconds, functioning as a "Shazam for car diagnostics."
- A natural language interface replaced physical manuals, allowing mechanics to retrieve remediation steps instantly.
- The initiative promises millions in savings across thousands of dealerships and multiple car models.
- Case Study: Pharmaceutical Industry
- Novo Nordisk utilized Large Language Models (LLMs) trained on approved drug data to draft Clinical Study Reports (CSRs).
- The system reduces initial drafting cycles from manual effort to mere minutes.
- The AI-generated drafts demonstrate higher quality than typical manually produced versions.
- The use case demonstrates immediate ROI in heavily regulated, high-scrutiny industries.
- Role of the Database Layer
- Databases are evolving from simple record-keeping to becoming the "memory" and state layer for AI applications.
- Unstructured data (70% of global data) is being made actionable through embeddings and probabilistic software.
- Advanced enterprises are moving beyond simple chat interfaces to complex RAG workflows that merge metadata, vector embeddings, and transactional data.
- High-quality retrieval requires filtering unstructured data based on metadata and keyword relevance, not just semantic similarity.
- Vector Databases and MongoDB's Strategy
- MongoDB's vector capabilities originated from semantic search needs in e-commerce before the generative AI boom.
- The goal is to eliminate the need for "RAG gymnastics" by keeping vector search, metadata, and transactional data in a single system.
- Vector databases are viewed as a foundational primitive (like B-tree indexes) rather than a temporary trend or a replacement for core databases.
- Vector search, graph structures, and metadata filtering are complementary modalities used together for high-quality retrieval.
- AI Transformation as Business Transformation
- MongoDB's shift from on-prem to cloud was treated as a holistic business transformation, not just a new product launch.
- Every functional leader was integrated into the strategy, with incentives, financial models, and customer success processes aligned to the cloud-first model.
- Top-down support and "shuttle diplomacy" were critical to overcome organizational inertia and align sales teams with the new consumption model.
- The strategy avoids forcing customers into the cloud, instead focusing on enabling customer choice through integrated PLG and Enterprise sales.
- Developer and Agent Experience
- Developer Experience (DX) remains the primary driver of adoption, but the focus is shifting toward "Agent Developer Experience."
- As software becomes AI-generated, the database must provide a compelling interface for trillions of agentic developers.
- MongoDB aims to be the default choice for agents by managing structured, semi-structured, and unstructured data effectively.
- The company is collaborating with AI labs to define "canonical training data" and quality standards for MongoDB code generation.
- Future Hypotheses and Outlook
- LLMs will act as the "brain" for reasoning, while databases act as the "memory" for state and world state.
- Foundational model APIs and databases are complementary side-by-side components, not a direct substitute for one another.
- The next trillion tokens of training data are expected to come from private enterprise data, primarily used for RAG and fine-tuning.
- MongoDB does not claim ownership of customer data for external model training; data remains with the customer.
- Ben Thompson's theory suggests the industry is currently focused on extracting meaning from vectors rather than just storing them.
- Rapid-Fire Insights
- Favorite AI Apps: Google Gemini Deep Research, Perplexity, and Snipped (for podcast summarization).
- Professional Admiration: Inspired by researchers who see the future trajectory and industry leaders navigating the "battle of titans."
- Developer Future: Agrees that "every developer will become an AI engineer" as generative AI democratizes capabilities beyond centralized data science teams.