The Data Architecture Behind Xapa’s AI and the Company Vault (Part 2 of 3)

By Leon Papkoff, CTO, Xapa

In Part 1 of this series, we introduced Context Engineering, our approach to making sure Xapa AI gets the right information at the right time so it can coach like a real human would. If you missed it, the short version is this: more data doesn't make AI smarter. Better data does.

Now in Part 2, I want to take you behind the curtain and show you how we supercharged that context engine by plugging in something we've never had before, your company's own knowledge. And I want to be honest about the engineering challenges we hit along the way, because the breakthroughs only make sense if you understand the problems we had to solve first.

Here's what I'm most excited to share, internal accuracy jumped from ~70% in Xapa AI 1.0 to over 94% in 2.0. And that number didn't come from throwing more data at the problem, it came from fundamentally rethinking how our AI reasons about the data it already has. I'll walk you through exactly how we got there, what broke along the way, and why the setbacks ended up being the most important part of the story.

With the launch of Xapa AI 2.0, we restructured the core of the AI, expanded its reasoning capabilities, and, most importantly, gave it an entirely new source of intelligence: the documents, policies, and institutional knowledge that make your company, your company.

Meet Company Vault

At the heart of this upgrade is Company Vault, and I genuinely believe it changes what's possible with enterprise AI coaching.

Here's the idea in plain terms. Before Company Vault, Xapa AI was already pretty smart. It knew you – your role, your communication style, your goals. And it knew soft skills – how to give feedback, how to lead meetings, how to navigate conflict. But it didn't know your company. It couldn't tell you what your vacation policy says, or walk you through your organization's specific sales methodology, or reference your onboarding process by name.

Company Vault changes that. For the first time, Xapa AI doesn't just understand people and soft skills, it understands your enterprise.

With Company Vault, administrators upload company-specific materials into a secure, AI-optimized repository. Once those files are processed (I'll explain how in a minute), they become part of Xapa AI's reasoning fabric, meaning the AI can actually draw on your company's real policies, culture, and workflows when it coaches your employees. Xapa AI speaks your language now, grounded in your knowledge.

A quick note on security, because I know this is top of mind for any enterprise team evaluating a platform like ours. Company Vault was built with enterprise-grade security from day one. All uploaded content is encrypted at rest and in transit. Your data is fully tenant-isolated, no other customer can ever access your documents, and role-based access controls ensure that only authorized users see the information they're supposed to see. Your Company Vault data is never used to train, fine-tune, or otherwise improve any underlying language model, not ours, not our LLM partners'. Your content is retrieved at query time to inform a response, and that's it. It stays yours. We take data stewardship seriously because our enterprise customers expect nothing less.

Company Vault supports a wide range of internal resources that can be uploaded and adjusted at any time:

• Mission, vision, and values: The foundational statements that define your organization's purpose and culture
• Policies, handbooks, and compliance docs: Employee manuals, code of conduct, legal guidelines, the rules everyone needs to follow
• Playbooks, SOPs, and onboarding guides: Step-by-step procedures for how work actually gets done, department by department
• HR calendars and resources: Holiday schedules, benefit guides, wellness program info
• Sales methodologies, product docs, and training materials: Your sales process/framework (MEDDIC, Challenger, whatever you use), product FAQs, training decks
• Everyday operational info: Cafeteria menus, IT support procedures, office logistics, the little things that keep everything running

Here's the key part, each file you upload is automatically chunked, indexed, and vectorized (I promise I'll explain what all of that means shortly). The result is that your content becomes instantly searchable, not in a basic keyword way, but in a way where the AI actually understands what the content means. It's the difference between a search engine and a colleague who's read every document in your company and remembers all of it.

Why This Required Us to Rethink Everything

I wish I could tell you that adding Company Vault was like plugging in a new feature. It wasn't. It forced us to rethink how our entire AI model handles context.

Here's why. Company documents are dense. They're varied. They're interconnected in ways that aren't always obvious. And there can be a lot of them. If we had simply dropped all of that content into Xapa AI's existing architecture, the AI would have drowned in information, which is exactly the "context overload" problem we talked about in Part 1. You can't just dump an entire intranet into an AI and expect it to behave intelligently. Trust me, we tried early on. It got confused, not smarter.

So we made the hard (but necessary) decision, rebuild the way Xapa AI reasons from the ground up. We extended our Context Engineering discipline into enterprise data itself by building:

• Advanced vector databases for high-speed, meaning-based search across large volumes of text
• Smart chunking with labeled embeddings – breaking documents into bite-sized, tagged pieces (explained below)
• Semantic + keyword hybrid search – combining meaning-based search with exact keyword matching
• A redesigned retrieval pipeline (Outer-Loop & Inner-Loop) built specifically for precision, relevance, accuracy and efficiency

Think of it this way, we taught Xapa AI to treat your company's knowledge the way a world-class librarian treats an archive, everything is meticulously indexed, contextually organized, and surfaced only when it's actually needed.

Three Knowledge Sources, One Unified Intelligence

To build Company Vault, we first had to step back and map out the different types of knowledge that Xapa AI draws from. It came down to three primary sources, what we call our three "tool sets":

1. Employee Directory (People Context): Think of this as a living org chart on steroids. It's a digital directory of everyone in your organization, names, titles, departments, reporting structure, plus work-style profiles, areas of expertise, and communication preferences (if you choose to include them). This gives Xapa AI social and organizational awareness. It knows who works with whom, and how.

2. Xapa Knowledge Center (Soft-Skills Content): This is the library we've always had curated leadership and soft-skill training developed by coaches and behavioral scientists. Communication, conflict resolution, feedback, emotional intelligence, and more. It's the "people skills" brain that powers our coaching.

3. Company Vault (Institutional Knowledge): The new kid on the block. Your company's own documents, handbooks, policies, calendars, playbooks, product FAQs, onboarding guides, SOPs. Everything that captures how your organization operates and what it values.

In Xapa AI 1.0, the AI leaned on the first two sources. The game-changer in 2.0 is adding that third source, your Company Vault, into the mix. Now Xapa AI can weave together personal context, organizational context, and expert coaching guidance in a single response. That's incredibly powerful.

But making three very different data sets play nicely together? That required some serious innovation.

Vectorization: Teaching the AI to Search by Meaning

Let's talk about vectorization, because it's one of those technical concepts that sounds intimidating but is actually pretty intuitive once you see what it does.

Normally, when you search for something, the system looks for exact keyword matches, you type "vacation policy" and it finds documents that contain those exact words. That works fine... until someone phrases their question differently. What if they ask "Can I roll over my unused PTO?" That's the same question, but the words don't match.

Vectorization solves this. It converts text into numerical representations, essentially translating words into math, in a way that captures the meaning behind the text, not just the literal words. Once text is vectorized, the AI can perform semantic search, finding information based on what it means, not just what it literally says.

We built a separate vector index for each of our three knowledge sources:

• People Vector Index: Employee profiles converted into vectors. Ask "Who in sales has SaaS experience?" and Xapa finds the right person by meaning, even if the profile says "enterprise software sales" instead of "SaaS."
• Content Vector Index: Our soft-skill training content, chunked and vectorized. The AI finds relevant coaching guidance even when the user's phrasing doesn't match a lesson's exact wording.
• Document Vector Index: Every company document, sliced into pieces and embedded as vectors. A 90-page employee handbook might become dozens of vectors, one for the vacation policy section, another for code of conduct, another for dress code, each capturing that snippet's meaning.

Here's a real example of why this matters: An employee asks, "Can I carry over vacation days to next year?" Our vector search pulls up a chunk from the handbook about "unused PTO carryover" and another from an FAQ about "vacation rollover policy", even though the employee never used those exact words. The vectors bridge the gap between how people naturally ask questions and how documents are actually written.

Smart Chunking: The Art of Breaking Things Down

A crucial part of vectorization is deciding how to break documents apart. We don't take an entire 90-page handbook and turn it into one massive vector, that would be too bloated and vague for the AI to do anything useful with. Instead, we split documents into what we call semantically coherent chunks, small sections (usually a paragraph, sometimes a sentence or two) that hold together as a complete thought.

Here's why this matters in practice. If a Sales Playbook has a section called "Qualifying Leads" and another called "Closing Deals," we treat those as separate chunks. So when you ask about qualifying leads, Xapa retrieves just that specific snippet, not the entire 50-page manual. You get precision instead of noise.

We also tagged each chunk with metadata, basically a digital sticky note that says where it came from. Something like: "Source: Employee Handbook, Section: Vacation Policy, Last Updated: 2024."

These tags do two important things:

Focused Retrieval: They let us filter searches intelligently. If someone asks about the holiday calendar, we can restrict the search to just chunks tagged as "Holiday Calendar" rather than searching through every document. This dramatically improves accuracy.

Context and Credibility: They tell both the AI and the user where the information originated. When Xapa AI gives you an answer, it can say "according to the Company Handbook, Vacation Policy", and suddenly you're not just trusting a chatbot, you're trusting an AI that shows its sources. That transparency is huge for building confidence with users on the platform.

Hybrid Search: Best of Both Worlds

Even with semantic vectors doing their thing, we didn't abandon traditional keyword search. Instead, we built a hybrid search approach that combines both, because sometimes you need the brains of meaning-based search and the precision of exact matching.

Here's a scenario. Someone asks: "What does the ABC Initiative document say about data privacy?" The term "ABC Initiative" is very specific. Semantic search will catch the general idea, but we also want to directly find that exact phrase.

So when Xapa AI queries the Company Vault, it actually runs two searches in parallel:

• Semantic search; Finds content that means the same thing as the question
• Keyword search: Catches exact matches, names, and rare terms

Then it blends the results. Searching for "probation period policy" semantically finds relevant chunks even if the document says "introductory period," while keyword search directly hits on "probation" if it appears anywhere. You get higher recall (finding more of what's relevant) and high precision (not burying the answer in noise).

One Question, Three Parallel Retrieval Paths

Here's something that anyone who works in an office knows intuitively: people don't ask simple, one-dimensional questions. They ask layered questions that combine multiple intents, exactly the way you'd ask a knowledgeable colleague.

For example, someone might say:

"Help me onboard a new sales hire by outlining the onboarding steps, summarizing our sales process, and identifying the best mentor to pair them with."

That's three questions hiding inside one sentence. And in Xapa AI 1.0 (like most AI systems), all three intents got crammed through a single retrieval funnel. The result? Answers that were technically correct but diluted, generic, and disconnected from how the company actually operates.

Xapa AI 2.0 handles this completely differently.

Every user prompt is treated first as a reasoning problem, not just a search query. The system analyzes the request, identifies the distinct intents, and fans out into three parallel retrieval paths, each handled by a specialized sub-agent:

Path 1 – Knowledge Center Agent: Pulls best-practice onboarding guidance from the Xapa Knowledge Center, checklists, learning paths, early milestones for ramping up a new hire.

Path 2 – Company Vault Agent: Retrieves your company's specific sales methodology using hybrid search. Maybe that's your MEDDIC playbook, your CRM process docs, your deal review templates. Not generic sales advice, your sales framework.

Path 3 – Employee Directory Agent: Searches people profiles to identify ideal mentors based on role, tenure, expertise, and mentorship track record. It might suggest "Jordan Lee, Senior Account Executive, has onboarded multiple new hires and has a strong mentorship record."

Once these three sub-agents finish their work, the system merges and re-ranks the findings, scoring for relevance, authority, and signal strength, and discarding anything that's low-value or redundant. Only the highest-signal, most relevant information makes it into the final prompt.

Then Xapa AI reasons across all of it to compose a single, coherent response. The user never has to think about which system handles which sub-question, they just receive an answer that feels complete, personalized, and grounded in their organization's reality.

The Efficiency Breakthroughs That Brought It All Together

Okay, here's the part I haven't told you yet, and honestly, it's the part I'm most proud of from an engineering standpoint.

When we first added Company Vault data into our existing pipeline, Xapa AI actually got more confused, not less. More context was degrading performance, not improving it.

That was a sobering moment. We had all this incredible enterprise data flowing in, and the AI was struggling to make sense of it. The potential was there, but the execution wasn't landing. We needed to get surgical about efficiency at every layer of the system.

That was a sobering moment. We had all this incredible enterprise data flowing in, and the AI was struggling to make sense of it. The potential was there, but the execution wasn't landing. We needed to get surgical about efficiency at every layer of the system.

That's when we introduced what we internally call our Outer-Loop Efficiencies and Inner-Loop Efficiencies, and they changed everything. If Context Engineering is the strategy, these are the execution systems that make it real. Together, they're what took Xapa AI from "promising but overwhelmed" to "fast, focused, and accurate."

Outer-Loop Efficiencies: Making Smarter Decisions Before We Even Search

The outer loop is about preparation, all the decisions the system makes before it goes looking for information. Think of it like a great researcher who plans their approach before diving into the library, rather than randomly pulling books off every shelf.

Here's what that looks like in practice:

• Staying focused and direct: We built explicit rules that keep responses clean and concise. No restating your question back to you, no rambling preambles, no unnecessary filler. One clear, useful thought at a time. It sounds like a small thing, but it had an outsized impact on the quality of every response.
• Asking before assuming. Rather than making its best guess and searching in the wrong direction, Xapa AI will pause and ask you a quick clarifying question when your intent isn't entirely clear. This prevents a cascade of wasted searches and keeps accuracy high.
• Raising the quality bar before retrieval even begins: We set strict confidence thresholds across every knowledge source. Generic documents with no real bearing on your question get excluded upfront, and content that doesn't score high enough as a relevant match simply doesn't make the cut. No borderline, "close enough" results sneaking into your response and muddying the coaching. The filtering happens before the search, not after.
• Using only the tools that are actually needed. Not every question requires every data source. The system is smart enough to selectively engage only the retrieval paths that are genuinely relevant to your query and skip the ones that aren't. Less noise, faster answers.

Inner-Loop Efficiencies: Making the Engine Leaner and Faster

If the outer loop is about smart preparation, the inner loop is about efficient execution, what happens under the hood while the system is actively retrieving and assembling information. This is the engineering work that users never see, but absolutely feel.

Here's what's happening behind the scenes:

• Trimming the fat on content: AI models process information in units called tokens, essentially small chunks of text. The more tokens a model has to process, the slower and costlier the response. We apply careful trimming to every piece of content that enters the system, cutting it down to only what's essential. The result is faster responses without sacrificing quality.
• Caching and reusing data intelligently: Rather than making repeated round trips to the database for the same information, we cache frequently used data and reuse it across a single request. It's the difference between a librarian who walks to the back room once to get everything you need, versus making ten separate trips for each item.
• Parallelizing retrieval: Instead of handling one piece of information at a time, the system batches multiple retrievals, all three sub-agents working simultaneously, into a single operation wherever possible. At enterprise scale, this kind of parallel processing is the difference between a system that feels instant and one that feels like it's thinking a little too hard.
• Loading resources on demand: Rather than pre-loading every tool and resource at startup "just in case," the system loads what it needs only when it's actually required for a specific query. It's a behind-the-scenes efficiency that keeps the system lean and responsive at all times.

These inner- and outer-loop optimizations are the unsung heroes of Xapa AI 2.0. They're what transformed Company Vault from a theoretical breakthrough into something that actually works at speed and scale. Without them, the Company Vault would have remained an interesting idea buried under its own weight. With them, everything clicks into place, and the experience on the other end feels less like interacting with a database, and more like talking to someone who already knows exactly what you need.

Less Context = A More Accurate (and More Useful) Xapa AI

After all these steps, vectorizing, chunking, searching, filtering, optimizing, and merging, what actually reaches the AI model is a dramatically smaller and higher-quality bundle of information than what we started with. We might begin with thousands of potential data points spread across your company's knowledge, but by the time we assemble the final prompt, we're down to maybe a half-dozen snippets that are exactly what's needed.

This reduction isn't a technical footnote, it's the breakthrough. By stripping away noise and eliminating what we call "context rot" (irrelevant, stale, or redundant information that dilutes the AI's reasoning), we taught Xapa AI to stop trying to read your entire intranet and instead focus only on the data that matters for this specific question, for this specific person, right now.

The numbers speak for themselves: internal accuracy jumped from ~70% in Xapa AI 1.0 to over 94% in Xapa AI 2.0 during testing.

How We Know It Works

I mentioned the accuracy jump earlier, ~70% to over 94%, and I want to be transparent about what that actually means, because a number without methodology is just marketing.

We built a dedicated evaluation console in our development environment specifically designed to stress-test every layer of the system. It lets us generate hundreds of test scenarios, real-world queries spanning all three knowledge sources, and measure end-to-end answer correctness, did the AI retrieve the right information, from the right source, and deliver a response that actually answers the question accurately and completely?

We're not just measuring one thing. The console tracks retrieval precision (did we pull the right chunks?), response accuracy (is the final answer factually correct given the source material?), latency (how fast did we get there?), and coherence (does the response read like a well-reasoned answer or a pile of stitched-together fragments?). Every adjustment we make, whether it's tuning our outer-loop filtering, tightening confidence thresholds, refining chunking strategies, or even swapping underlying LLMs as new models hit the market, gets benchmarked against these metrics before anything ships to production.

That last point matters more than people realize. The LLM landscape is evolving fast, and our evaluation framework gives us the ability to test new models, new tool sets, and new retrieval strategies in isolation, measuring their impact on accuracy and performance before a single customer ever sees the change. It's how we future-proof the system.

The ~70% to 94% jump reflects end-to-end answer correctness across a broad set of representative queries, the kind of layered, multi-intent questions real employees actually ask. Not cherry-picked demos. Not simple factual lookups. The messy, complicated stuff.

When the model isn't spending its processing power on irrelevant details, it gives dramatically better answers. It's the difference between dumping a truckload of documents on someone's desk versus handing them a laser-focused briefing tailored to their immediate question.

Under-the-Hood Benefits

Beyond accuracy, this re-architecture sets Xapa up for long-term success in ways that matter to every organization we work with:

• Scalability: Each vector index scales independently. If your company doubles in size, the people index grows, but document searches don't get slower. Upload hundreds of new files? The employee directory stays fast. Each part of the knowledge base expands without bogging down the others.
• Maintainability: Discrete chunks with metadata make updates straightforward. Change a policy? We re-index just that section. Remove a document? Those vectors get dropped. And since every snippet knows its source and timestamp, the system automatically favors the latest information, so Xapa AI won't accidentally give you last year's policy when a new one exists.
• Transparency & Trust: Because the AI retrieves actual snippets from official, trusted sources, answers can include references like "according to the 2024 Employee Benefits Guide." Users can see why Xapa answered the way it did. That kind of transparency makes the difference between an AI people tolerate and an AI people actually trust.
• Flexibility: The framework is modular by design. Today we have three knowledge sources, but tomorrow we could plug in more, access to your work calendar, a live data feed from your business systems, without overhauling the architecture. Xapa AI can evolve as your needs grow.

The Results: What This Means for People Who Use Xapa Every Day

Let me step back from the technical details for a moment, because at the end of the day, architecture only matters if it makes a real difference for the people using the platform.

And it does. The difference is night and day.

The AI's context window has shrunk dramatically, it's processing far less text per query, but the quality of its output has gone through the roof. Token usage has plummeted, which means Xapa AI responds faster and at lower computational cost. Precision has soared, the AI consistently locks onto the most relevant information and ignores the noise. And recall has strengthened too, it misses less, which means responses are more comprehensive and complete.

But here's what really matters: what this feels like for someone learning and growing on Xapa every day.

Imagine you're a new manager preparing for your first tough feedback conversation. In the old system, you might get a solid but generic tip about feedback techniques. Now, Xapa AI 2.0 knows your communication style is "Driver" (direct, results-oriented), and it knows your team member is a "Stabilizer" who values harmony. It pulls a relevant coaching lesson from the Knowledge Center, checks your company's feedback guidelines from the Company Vault, and tailors the advice to both of your personalities. The guidance you get isn't theoretical. It's grounded in who you are, who you're talking to, and how your company expects these conversations to happen.

Or imagine you're onboarding and trying to figure out how things work at your new company. Instead of digging through a SharePoint folder or pinging five different people on Slack, you ask Xapa AI 2.0. It pulls the answer from the actual onboarding guide your company uploaded, cites the source so you know it's trustworthy, and connects you to a relevant Xperience so you can build the skills to succeed in your new role. All in one conversation.

That's the real promise of this re-architecture. Xapa AI's answers now feel on-point, personal, and immediately useful. You ask a question, and the response references the specific policy, the specific person, the specific training lesson that matters, not a bunch of generic fluff. It feels less like talking to a chatbot and more like getting advice from a colleague who knows your company inside and out and happens to be an expert in leadership and professional development.

This is exactly why we built Xapa AI 2.0, to ensure every coaching moment feels relevant, trustworthy, and grounded in the real world your employees navigate every day.

What's Next

In Part 3 (the finale of this series), we'll bring everything together and show how it comes alive in real-world scenarios. We'll highlight high-impact use cases where personal context, company documents, and soft-skill knowledge weave together into powerful, enterprise-ready solutions that drive growth and productivity.

Stay tuned, Part 3 will show you what all this innovation actually does for your organization in practice.

By Leon Papkoff, CTO, Xapa