Skip to content

Private betaTyped product knowledge for teams building with AI

Your specs are data. Treat them like it.

Spacia turns PRDs, specs, decisions and tests into a typed, versioned knowledge graph — so people and AI agents can ask structural questions and get exact answers, not guesses from prose.

See how it works
  • Your models, your keys
  • No AI markup
  • Self-hostable
PRDRefunds v2v3

1.Business rules

BR-12Customers can request a full refund within 14 days of purchase.

2.Requirements

FR-31Self-serve refund button in billing settings.

FR-32Check eligibility before a refund is issued.

3.Acceptance

TC-89Refund is blocked on day 15.

Ask the graphWhat implements BR-12?
FR-31, FR-32 · 2 resultsIllustrative

01 — The problem

Your knowledge lives in prose. Your agents have to guess.

A PRD in Notion. A decision in Confluence. Rules in CLAUDE.md, AGENTS.md and .cursorrules. The real answer in a Slack thread. Ask an agent what constrains your refund flow, and watch what it actually does.

CLAUDE.md, AGENTS.md, .cursorrules, PRD-refunds-v3-FINAL.docx, Notion › Refund policy, Confluence › Payments, docs/adr/0042-stripe.md, #billing — “wait, is it 14 or 30 days?”

What business rules constrain the refund flow?

  1. ~ grep -ri "refund" docs/probabilistic
  2. ~ glob **/*payment*.mdprobabilistic
  3. ~ read PRD-refunds-v3-FINAL.docxprobabilistic
  4. ~ follow [[refund-policy]]probabilistic
  5. ~ synthesize an answerprobabilistic

“Refunds are allowed within 14 days.”

3 sources read · 2 disagree · 1 never found

02 — The diagnosis

We’re asking models to do structural analysis on unstructured data.

  1. 01

    Scattered

    Every team keeps its own copy, and the copies diverge.

  2. 02

    Disconnected

    An edit in one place never reaches the others.

  3. 03

    Unaddressable

    A rule is a sentence, not an object you can point at.

  4. 04

    Non-deterministic

    Every answer is a heuristic. None of them is a guarantee.

+20%
Repository context files like AGENTS.md “do not generally improve task success rates”, while raising inference cost by over 20% on average.
Gloaguen et al., Evaluating AGENTS.md, arXiv:2602.11988 (2026)
45%
of developers say their top frustration with AI is answers that are “almost right, but not quite.”
Stack Overflow Developer Survey, 2025

03 — The inversion

So we inverted it.

Documents stay for people. Underneath, every rule, requirement, test and decision becomes a typed object — with an id, a version and explicit relationships.

  1. 01

    Documents are written for people.

    Prose is great to read and terrible to traverse.

  2. 02

    Spacia breaks them into typed objects.

    Rules, requirements, tests, APIs and decisions, each with an id and a version.

  3. 03

    Duplicates merge. Conflicts surface.

    A reconciler checks every draft against what already exists. Nothing is silently overwritten.

  4. 04

    Relationships become edges.

    Explicit, directional, typed, and pinned to the exact version they point at.

Every team builds this chain. Spacia models it.

04 — The proof

One rule changes. Eight artifacts drift.

Your refund window moves from 14 to 30 days. In a folder of documents, nobody knows what just broke. In Spacia, it’s a lookup.

  • Current
  • Stale — pinned to v1
  • At risk — downstream

4 stale · 4 at risk

IllustrativeStale flags on version-pinned links ship today; automated downstream analysis is on the roadmap.

“Which artifacts are out of date with the refund rule?”

Probabilistic

An agent reads the PRD, the help article and a few tests, matches them by wording, and estimates which ones changed. It answers confidently. 

Depends on what it happened to read. It might miss TC-89.

DeterministicIllustrative query

FOR edge IN edges(to: BR-12)  WHERE edge.pinned < BR-12.version  → STALE edge.fromWALK downstream(STALE)  → AT RISK

STALEFR-31 · FR-32 · DOC-3 · DOC-5

AT RISKTC-88 · TC-89 · API-7 · TASK-204

Milliseconds. Exhaustive. Zero tokens.

05 — The product

Familiar on the surface. A graph underneath.

  1. 01Model

    Define what your knowledge is.

    Document types, artifact types and the relations between them — or start from a template: PRD + MRD, Tech Spec + ADR, Discovery & Research, Incident & Postmortem, Strategy & OKRs.

  2. 02Write

    Write like a doc. Every mention is an edge.

    A block editor your team already knows how to use. Mention a rule and you’ve created a typed, versioned link.

  3. 03Ask

    Or describe it, and let the agent draft.

    Chat with an agent that reads your schema, searches the graph and drafts structured changes — using your own model keys.

  4. 04Review

    Every change is a draft you approve.

    The reconciler reuses what exists, links what’s related and flags what conflicts. You apply it in one atomic step.

  5. 05Trace

    Every version. Every edge. Forever.

    Versions are immutable. Edges remember the version they meant. Stale links surface on their own.

06 — Under the hood

Built to stay true as you change.

  • 01

    Constructive-only AI

    The agent can propose new types, relations and fields. It can’t delete anything — destructive operations are unrepresentable.

  • 02

    Brownfield import

    Bring the folder of markdown you already have. Review every file in a sandbox, promote it atomically, roll back in a click.

  • 03

    Project constitution

    Write your premise once — audience, tone, non-negotiables. It rides along on every AI workflow.

  • 04

    Context cards

    Glossaries, personas, policies. The agent sees what’s available and pulls in what it needs.

  • 05

    Hybrid search

    Vector and keyword search over every node, so the agent finds what exists before it writes something new.

  • 06

    Bring your own model

    Anthropic, OpenAI, OpenRouter — or local models through LM Studio. Your keys, your bill, no markup.

  • 07

    Templates

    Start from PRD + MRD, Tech Spec + ADR, Discovery & Research, Incident & Postmortem or Strategy & OKRs — or from a blank schema.

  • 08

    Export everything

    One archive holds your schema and graph, plus your own chats and plans. Import it into any Spacia instance.

  • 09

    Teams and orgs

    Personal and organization projects with members and roles. Projects you can’t access simply don’t exist for you.

07 — Trust

AI that can’t corrupt your knowledge.

The agent never writes to the graph. It proposes typed commands; a deterministic writer validates them and commits each draft in a single transaction.

Reasoning: Agent. Truth: Validator → Graph. create BR-12 ✓ · relate FR-31 → BR-12 ✓ · update PRD-14 v3 → v4 ✓ · delete PRD-14 ✕ unrepresentable.

The model decides. The writer commits.

LangGraph proposes · Go validates · SurrealDB commits

  • Typed commands, never direct writes
  • Human review on every draft
  • Sandboxed imports, one-click rollback
  • Immutable versions
  • Credentials encrypted at rest
  • Self-hostable on your own infrastructure

08 — For your team

Questions you can finally answer.

Which requirements still assume the old pricing rule?

  1. FR-14Prorated downgrade refundimplementsBR-7v1Stale
  2. FR-22Annual-plan refund creditimplementsBR-7v1Stale
  3. FR-40Invoice line totalsreferencesBR-7v1Stale
Illustrative

09 — What’s next

Where the graph goes next.

Shipped

  • Typed schema and templates
  • Block editor with mention edges
  • AI drafting with reconciliation
  • Immutable versioning
  • Brownfield import with rollback
  • Constructive-only schema drafts

Next

  • MCP server for coding agents
  • Blast-radius analysis before a change lands
  • Approval as a dialogue, not a gate

Later

  • Connectors for Notion, Confluence and Linear
  • Reactive nodes that compute from their neighbours

Roadmap items are plans, not promises.

10 — Pricing

Free during the private beta.

Beta teams get everything free. Here’s what pricing will look like at launch — with your keys, no AI markup and no credits.

Community

For individuals and evaluation

$0

  • Self-host with Docker
  • Bring your own model keys
  • Every core feature
  • Cloud: up to 3 members and 2 projects

Team

Most teams

For product and engineering teams

$20

per editor / month

  • Unlimited projects
  • Organizations and roles
  • Brownfield import
  • Export and import
  • Priority support
  • Viewers and reviewers are free

Enterprise

For security-sensitive organizations

Custom

  • SAML SSO and SCIM
  • Audit logs
  • Supported or air-gapped self-hosting
  • SLA and onboarding

Planned pricing — subject to change before general availability.

11 — FAQ

Questions, answered.

Is Spacia a Notion or Confluence replacement?

For specs and decisions, yes. You still get documents your team can read and write, but underneath, every rule, requirement and test is a typed object. Meeting notes and wikis can stay where they are.

How is this different from RAG or a vector database?

RAG finds text that looks similar to your question. Spacia stores what your knowledge is and how it connects, so “what’s downstream of this rule?” is a traversal with an exact answer. Spacia still uses hybrid search to find things — it just doesn’t stop there.

Which models does it work with?

Anthropic, OpenAI and OpenRouter, or local models through LM Studio. You bring your own keys and choose models per project.

Can the AI delete or overwrite our work?

No. Graph changes arrive as drafts you review and apply. Schema proposals are constructive-only — the payload has no delete operation at all. Published versions are immutable.

What happens to the documents we already have?

Import a folder of markdown. Each file becomes a draft you review in a sandbox; when you approve the batch it’s promoted atomically, and the previous state is archived so you can roll back.

Can Claude Code, Cursor or Codex use it?

Inside Spacia, the agent already queries the graph. An MCP server, so external coding agents can do the same, is next on our roadmap.

Can we self-host?

Yes. Spacia runs on Postgres and SurrealDB with Docker Compose, and every project exports to a single archive you can import elsewhere.

Is our data used to train models?

Spacia doesn’t train models. Model calls go to the provider you configure, under your own account and its terms.

Start treating your specs like data.

Spacia is in private beta. Request access, and we’ll help your team import what it has and ask its first structural question.