“95% of our retrieval is done by AI agents.” - Sergio Prada, Co-Founder & CTO, Metal
Metal needed a retrieval foundation that could evolve as fast as their product, without hitting a wall.
Introduction
Private equity firms manage vast amounts of unstructured data, including deal documents, expert call transcripts, financial statements, CRM records, and more. The challenge isn’t simply accessing this information. It’s connecting and understanding it, in context, across the investment lifecycle.
Metal AI was built to address this challenge. Its purpose-built institutional intelligence platform, used by established private equity firms transforms fragmented historical and live deal data into a living system of record that drives conviction at every stage of the investment lifecycle.
To deliver this vision at scale, Metal leverages Vespa.ai as its core retrieval layer, powering entity relationships, advanced ranking, and real-time context-aware retrieval across complex investment data.
The Need for Relationship-Driven Retrieval
As Metal’s product evolved, the limitations of traditional retrieval systems became clear.
Early architecture supported basic document search, but private equity workflows aren’t document-centric. They are entity- and relationship-driven. The enduring edge in private equity lies in drawing on decades of deal history, portfolio outcomes, and institutional knowledge. When that depth of experience surfaces reasoning and connections across time, every investment decision carries greater conviction.
Most traditional vector stores and search engines are fundamentally document-first. They index text, return similar passages, and rely primarily on semantic similarity or keyword matching. But for Metal’s use case, relevance requires more:
-
Understanding which answer is the most recent and legally approved
-
Identifying which company a metric belongs to
-
Connecting meetings to prior diligence activity
-
Applying business logic alongside semantic similarity
As Metal introduced more advanced workflows, like DDQ automation and agent-driven retrieval, the gap widened. Traditional systems struggle to:
-
Combine semantic similarity with recency and compliance rules within ranking
-
Support evolving data models without significant rework
-
Query across multiple object types in a unified way
-
Serve as a foundation for structured, iterative queries issued by AI agents
Layering custom logic on top of limited retrieval infrastructure would have created increasing technical debt, and each new entity type or ranking rule risked architectural compromise.
Metal needed a retrieval foundation that could evolve with the product, not constrain it.
Choosing a Retrieval Layer without Limits
Metal wasn’t simply selecting a search engine. They were selecting a long-term retrieval architecture.
Several capabilities distinguished Vespa:
-
Multi-entity modeling: Vespa supports multiple object types, like documents, people, activities, and financial data, as well as the relationships between them. This aligned with how Metal structures institutional knowledge.
-
Advanced ranking and filtering: Vespa can combine semantic similarity with structured filters like recency and business rules, enabling Metal to tailor retrieval to specific workflows.
-
Flexibility without re-architecture: New object types can be introduced without migrating existing data or rebuilding the system.
-
Operational simplicity: Moving to Vespa Cloud enabled the team to focus engineering capacity on product innovation instead of infrastructure.
These capabilities give Metal the ability to shape retrieval around business logic, rather than forcing business logic to adapt to infrastructure limitations.
“Our competitors focus on documents. With Vespa, we can focus on the full picture: companies, people, activities, and how they relate.” - Sergio Prada, Co-Founder & CTO, Metal
Architecture in Action
Metal treats retrieval as part of an AI agent orchestration layer, not just a standard search box.
When a user or agent asks a question like, “What’s this company’s EBITDA?”, the query is first interpreted by an AI agent. Rather than issuing a single plain-text search, the agent:
-
Determines which entity types to query (documents, companies, metrics, activities)
-
Applies structured parameters such as recency or workflow-specific filters
-
Executes retrieval against Vespa
-
Iterates as needed (paginating, refining, or querying related entities)
-
Assembles sufficient context before generating a response
Vespa powers this retrieval layer, enabling fast, structured queries across different object types and supporting the iterative retrieval process required by Metal’s agent-driven architecture.
Turning DDQ Chaos into Structured, Approved Intelligence
One clear example is Metal’s Due Diligence Questionnaire (DDQ) workflow. Private equity firms must respond to thousands of LP questionnaires using pre-approved answers. These responses cannot be freely generated by an LLM. They must come from content that has already been reviewed and approved by legal teams.
Answer banks change over time and are stored in unstructured formats like documents and spreadsheets. Metal indexes this data into Vespa, making the system aware of which documents are most recent. When answering a questionnaire, retrieval is prioritized not only by semantic similarity to the question but also by freshness.
This allows Metal to surface the most relevant and up-to-date approved answers, efficiently and reliably within its platform.
Scaling without Infrastructure Headaches
By building on Vespa Cloud, Metal achieved:
-
Improved feature velocity: The team can introduce new entity types and workflows quickly without architectural rework
-
Greater engineering focus: The team spends less time managing infrastructure and more time building differentiating product features
-
Scalable retrieval architecture: Metal can onboard new clients and data volumes without redesigning retrieval.
-
Confidence in long-term flexibility: Vespa is not a limiting factor as Metal expands into more advanced agent-driven workflows.
“Managing infrastructure can be a distraction. Vespa Cloud lets us focus on product.” - Sergio Prada, Co-Founder & CTO, Metal
Looking Forward: Build for an Agentic Future
Metal’s roadmap is deeply agentic. AI agents drive most interactions, deciding how best to query the platform and construct the context needed to answer sophisticated questions.
Because Vespa supports flexible, multi-entity retrieval with advanced ranking and real-time performance, Metal can:
-
Expand into more advanced analysis workflows
-
Build deeper relational structures between entities
-
Adapt retrieval strategies dynamically as business logic evolves
The result is an institutional intelligence platform that scales in both data volume and intelligence, evolving alongside the firm it serves.
“When you’re building something ambitious, you don’t want to hit a capability wall. Vespa gives us confidence that we won’t.” - Sergio Prada, Co-Founder & CTO, Metal
