What you can build
Pick one. Most of it already runs.
A gallery of what Vectros lets you ship on one secure back-end: each driven by the same typed records, documents, and hybrid search, reachable by your app or your agents, and each leading with something you can actually run. Three ship as bundled blueprints, one is a full reference app you fork and deploy yourself, and the rest are illustrative shapes the platform supports. We mark which is which on every card.
Agentic-SDLC Knowledge Base : your team’s engineering memory
Your decisions, conventions, runbooks, and gotchas as one searchable system of record: recalled by agents, browsable by your team.
The blueprint we run our own engineering org on. Model the whole SDLC knowledge lifecycle as nine typed schemas, split by content vs structure. ADRs, designs, references, runbooks, and post-mortems are documents (the prose is the artifact); controls, conventions, gotchas, and a glossary are records (the typed fields are). Two more tiers round it out: a private record each agent keeps for itself, isolated per principal, and a candidate queue holding unverified proposals pending review. The curated set links into a cross-surface knowledge graph and is recalled by hybrid search + grounded answers, so a cold-start agent asks “why is it shaped this way?” and gets the actual decision, cited.
What it exercises
- Documents and records, unified: prose artifacts (ADRs, runbooks, post-mortems) as documents; structured artifacts (controls, conventions, gotchas, glossary) as records: one index, one model.
- Cross-surface knowledge graph: typed references (a control points to the runbook that proves it, a convention to the decision that established it): navigable provenance, not just search.
- A private tier per agent: beyond the shared curated knowledge, each agent keeps its own private tier, isolated per principal by the platform: what it learns for itself, searchable and groundable by its owner alone.
- Hybrid search + grounded recall: recall by meaning, with citations: “why did we decide X?”, “have we hit this failure before?”
- Dual surface: agents capture and recall over MCP via a drop-in orientation prompt; your team browses the same typed context in the data-plane app.
Run it
vectros bootstrap --blueprint agentic-sdlc --no-seedBootstrap the agentic-sdlc blueprint (it ships seedless: you fill it from your own corpus), merge the scoped key into your MCP client, and point an agent at it with the bundled orientation prompt. This is the blueprint we dogfood our own engineering org on, paired with the separate Agent Memory package for Claude Code, which is designed around exactly this schema's memory and candidate-queue tiers. The most credible demo there is.
RAVV reference app : a full customer-facing app, zero backend of your own
Fork a full, production-grade app with its own sign-in, and the only thing you build is the front end.
An org manages compartments (cases, clients, accounts) for the people underneath it, and only the right people ever see any one of them. Sign-in happens through your own identity provider, its token exchanged directly from the browser for a scoped Vectros credential: no route of your own minting it, no backend of your own in between.
What it exercises
- Token exchange, not a credential route you run: your identity provider handles sign-in; its token is exchanged from the browser for a Vectros credential, scoped to exactly what that user is allowed to do.
- Per-org, per-client isolation: a case handler in one org cannot reach another org’s clients: enforced structurally by the platform, the same way it enforces isolation everywhere else.
Run it
This is a reference implementation, not a product to run as-is: the value is in the pattern, adapt it to your own domain and it holds exactly the same.
Second Brain : your personal knowledge base
Dump every note, idea, and link in one place, then just ask it.
Capture notes, ideas, bookmarks, and half-formed thoughts as typed records, then retrieve them by meaning instead of by remembering the exact words you used. Hybrid search spans keyword and semantic recall over a unified index, so “that article about pricing psychology” finds the note even when “psychology” never appears in it.
What it exercises
- Semantic + hybrid search: over your own content: keyword, semantic, or both in one call.
- Typed records: with lookups, so a note has structure (title, body, tags, source link), not just a blob.
- Grounded answers: ask a question and get a response cited back to the notes it came from.
Run it
vectros bootstrap --blueprint second-brainPick the second-brain blueprint, bootstrap a narrow key, then drive it from an MCP agent (“save this, summarize what I’ve captured about X”) or straight from the data-plane UI. Zero application code to start.
Chat with your docs : answers grounded in your own files
Upload your documents, ask in plain language, get answers that cite the source.
Ingest documents (inline or by file upload), organize them in folders, and ask questions across the whole set. Retrieval grounds the model on your own content, and every answer streams back with citations to the documents it drew from, so you can check the source instead of trusting a confident guess.
What it exercises
- Document ingest: text inline or file upload, organized in folders, with retrievable text and download URLs.
- Document/RAG retrieval: one search call spans documents and records together; no separate vector index to wire up beside your store.
- Citation-grounded answers: RAG over your data and document-scoped Q&A both stream with citations back to the source; a stateless chat mode is there for when you bring your own context.
Run it
This is an illustrative shape, not a bundled blueprint yet, but you can build it today on the same building blocks. Define a document-bearing schema, scaffold a blueprint from the bundled exemplars (vectros blueprint init my-docs --from task-management), bootstrap a key, and drive ingest and Q&A from an MCP agent or the data-plane UI.
Support & internal-wiki Q&A : one place your team and your customers can ask
Turn your help center and your internal wiki into something people can just ask.
Combine ingested documents (help articles, runbooks, policy pages) with typed records (products, plans, known issues) and let an assistant answer over both: grounded, cited, and isolated per customer when you run it for the customers you serve.
What it exercises
- Document/RAG over your knowledge base: articles and runbooks, retrieved by meaning.
- Facts and entities alongside the docs: structured records for products, statuses, and entitlements, searchable in the same call as the prose.
- Isolation, enforced automatically: when you run this as a product for your own customers, each customer’s context is a mandatory partition: a lookup that can’t prove it belongs there returns nothing, not a forgotten WHERE clause away from a cross-tenant leak.
Run it
Illustrative today. Model your articles as documents and your structured facts as records in a blueprint, bootstrap a scoped key, and drive question-answering from an MCP agent or the data-plane UI. When a customer later asks “is my data isolated?”, the answer is already built in: the same platform, no rewrite.
A small vertical SaaS : listings, search, and a helpful agent
Recruiting, tutoring, real-estate search: structured listings your users can search by meaning.
A worked example: a real-estate listing search (the data is synthetic). Model listings as typed records (price, beds, neighborhood, a free-text description) with lookups for exact and range filters and hybrid search over the description. A buyer asks “a quiet two-bed near good schools under $600k,” and the agent narrows by the structured fields and ranks the rest by meaning.
What it exercises
- Facts and entities: typed listing records with validated fields and lookupFields for exact, prefix, and range filters (price ranges, bed counts).
- Hybrid search: keyword + semantic over the free-text description, narrowed by the structured filters in the same query.
- Shared / multi-agent memory: context isolation plus scoped access, so the search agent and a follow-up agent share state without either one reaching outside its lane.
Run it
Illustrative: recruiting, tutoring, and listing search are all the same shape. Author a blueprint with your listing schema and lookups (vectros blueprint init listings), validate and plan it offline, bootstrap a key, then let an agent or your forked data-plane UI run the searches. All sample data synthetic.
Clinical Intake : structured intake with similar-case support
Structured intake that validates on the way in and surfaces the most similar prior cases by meaning.
The compliance peak of the gallery, and where the same platform that started lean shows what it grows into. Model an intake form as a validated record type, capture submissions (synthetic only), and use hybrid search to surface the most similar prior cases by meaning, as support for a clinician’s triage, never as automated decisioning.
This is triage support, not decisioning. It surfaces and structures information for a human; it does not diagnose, prioritize, or decide.
What it exercises
- Facts and entities with validation: the intake schema validates on the way in (required fields, formats, ranges), so structure is enforced at write time.
- Sensitive-field handling: fields you mark sensitive are destroyed before they reach retained history, masked on read unless a token is explicitly scoped to reveal them, and kept out of the search index entirely: three distinct, independent mechanisms.
- Tamper-evident audit and version history: every write to the audited intake model accrues a tamper-evident version record of what changed and who changed it, on a SHA-256 state-continuity chain that makes out-of-band alteration detectable.
- Isolation, enforced automatically: each context (a patient, a case, whatever your model calls it) is a mandatory partition derived from the credential, never a wildcard.
Run it
vectros bootstrap --blueprint clinical-intakeBootstrap the clinical-intake blueprint, then drive intake and similar-case lookups from an MCP agent or the data-plane UI. All example data is synthetic. The platform is a HIPAA-grade substrate. It does not, on its own, make your application HIPAA-grade.
The thread through them all
Different products, one back-end.
| Build | What it exercises | Today |
|---|---|---|
| Agentic-SDLC Knowledge Base | documents · records · graph · hybrid search · private memory | Bundled blueprint |
| RAVV reference app | BYO identity provider · zero backend · full customer-facing app | Reference app |
| Second Brain | semantic · facts/entities | Bundled blueprint |
| Chat with your docs | document/RAG · semantic | Illustrative |
| Support & internal-wiki Q&A | document/RAG · facts/entities | Illustrative |
| Small vertical SaaS | facts/entities · semantic · shared/multi-agent | Illustrative |
| Clinical Intake | facts/entities · semantic · document/RAG · shared/multi-agent | Bundled blueprint |
You declare the data model in one place and the platform gives you typed records, document ingest, and hybrid search over a unified index: not a database plus a vector index plus a sync pipeline plus an auth layer to wire together and operate forever. You pay for the capabilities you turn on. For what each capability costs, see the pricing page.
Start lean. Scale to compliance-grade without re-platforming.
Vectros is invite-only today. Self-serve and a free tier are on the way as we open access.