Knowledge-as-a-Service

Knowledge your agents can cite.

Turn scattered institutional knowledge into a retrievable, versioned graph. Point any LLM agent at your own endpoint and get answers grounded in your documents — with citations you can audit.

Your endpoint: http://acme.knowledge.servicev8.com/v1
Explore the graph
Draft documents clustering into the reviewed, graph-linked concepts your agents actually retrieve from.
Object storage for raw docs Git-curated knowledge pgvector recall Long-context expansion REST + MCP native
Why KnowledgeV8

Not an upload bucket. A knowledge system.

Most RAG products index everything and curate nothing. KnowledgeV8 treats your knowledge like source code — reviewed, versioned, and graph-linked — so answers stay trustworthy as the corpus grows.

Knowledge as code

Git-native curation

Author knowledge as plain markdown with frontmatter. Documents promote from draft to reviewed to authoritative as your team signs off — agents cite governed answers, not raw uploads.

Graph-aware retrieval

Whole concepts, not fragments

pgvector narrows a huge corpus to the few concepts that matter; long-context expansion then reads them whole — full body plus everything they link to — instead of stitching fragments.

Auditable answers

Every claim cites a source

Answers reference the concept ids they drew from. No black box — you can trace any statement back to the document it came from.

Model freedom

Bring your own model

Use your own OpenAI or Anthropic key at zero platform cost, or buy managed credits. No lock-in to a single provider's assistant.

Trust & access

Real access control, not a shared bucket.

Every agent, every document, every question — scoped, masked, and logged the way your own team already expects.

Flagship — permission-aware access

Agents see only what their human can

Connect Claude Code or any MCP agent to your workspace. Access is scoped live, per request, to the permissions of the person behind it — never a standing service account with broader reach.

No file archive

Processed, never held

Uploaded documents are parsed, indexed into concepts, and discarded. There's no growing pile of raw files sitting on our side for anyone to breach.

Masking, your call

PII stays out of the brain

Turn on masking before anything is stored — from fast pattern-based rules to LLM-assisted de-identification — so sensitive values never make it into a concept.

Full trail

Auditable by design

Every question is attributed to the person who asked it, every agent action is logged, and retention is a setting you control, not one we impose.

Stays current

The graph keeps working after you close the tab.

Sources sync themselves. Agents can contribute back. Nothing becomes truth without a person signing off.

Scheduled sync

Always fresh

Google Drive, Confluence, Zendesk, Git, S3, and more sync on a schedule, re-ingesting only what changed — never a stale answer from a doc that moved months ago.

Agent write-back

A brain that learns from your team's agents

Any agent can say "remember this" and deposit new knowledge mid-conversation — reviewable and attributed, but never authoritative until a human approves it.

Who it's for

One graph, every team's own doorway in.

Same endpoint, same citations, same curation lifecycle — each team just asks it different questions.

Engineering

Point your code agent at your own docs

Claude Code, Cursor, or any MCP client queries your runbooks, ADRs, and API references directly — no more grepping the wiki and hoping it's current.

connects via MCP — search_knowledge, read_concept, browse_index
Code agent, mid-task"How does the auth middleware handle token refresh?"
Tokens are refreshed 60s before expiry via a background interceptor; failed refreshes fall back to a full re-auth redirect rather than silently dropping the session. [playbooks/auth-token-refresh]
authoritative · reviewed by platform team
SRE / Ops

Cite the runbook, not your memory

Incident runbooks, postmortems, and escalation paths — retrieved and cited the moment an on-call engineer needs them, not rediscovered during the post-mortem.

connects via REST — POST /v1/answer, or MCP from a ChatOps bot
On-call engineer, 2am"What's the rollback procedure when the pgbouncer connection pool is exhausted?"
Drop connection_limit to 50% of baseline, restart pgbouncer, then watch for the queue depth alert to clear before restoring. Do not restart Postgres itself. [runbooks/pgbouncer-exhaustion]
authoritative · promoted after the last incident review
Sales

Answer the objection with the doc behind it

Battlecards, security questionnaires, and pricing policy — one query away, with a citation you can actually show the prospect.

connects via REST — POST /v1/answer from your sales-assist tool
AE, mid-call"Do we have SOC 2, and can I send the report?"
SOC 2 Type II, renewed annually; the report is under NDA — route the request through the security-review form, don't attach it directly. [sales/security-questionnaire]
authoritative · owned by security
Customer care

One policy, every channel, no drift

Refund windows, escalation policy, macros — support agents and your support bot pull from the same graph, so the answer doesn't depend on who's on shift.

connects via REST — POST /v1/answer from your helpdesk or support bot
Support agent"Customer's item arrived damaged, 22 days after purchase — refund or replacement?"
Damaged-on-arrival claims get a replacement or refund regardless of the 30-day window, no return required — photo evidence is enough. [playbooks/refunds]
authoritative · same policy your agents already cite
How it works

Three tiers in, a hybrid retrieval out.

Raw documents stay in your object storage. A small, git-curated layer holds the concepts worth trusting. pgvector is just the rebuildable index over both — recall, not the system of record.

1

Raw

Your S3 / GCS bucket or a git repo. Binaries and originals never leave your storage — we hold pointers.

2

Knowledge

Curated OKF concepts, reviewed and graph-linked. Small by design — PRs and diffs make sense here.

3

Recall

pgvector ANN narrows the corpus to the few concepts that matter. Recall only — it never sees the final answer.

4

Expand + answer

Full bodies plus one-hop linked neighbors go into context whole, then the model generates with citations.

Watch the full pipeline →
Agent-first

One endpoint. Any agent.

A REST API and a first-class MCP server. Claude, Gemini CLI, LangChain, or a bare curl loop — one config block and your knowledge is a tool.

Answers come back with a Citations: line listing the exact concept ids used — so you can verify, not just trust.
POST /v1/answer
# Ask your knowledge base a question
curl acme.knowledge.servicev8.com/v1/answer \
  -H "Authorization: Bearer kv8_…" \
  -d '{"question":"how do refunds work?"}'

# → grounded answer + auditable sources
{
  "answer": "Reverse the charge within the
            30-day window [playbooks/refunds].",
  "citations": ["playbooks/refunds"],
  "sources": [{ "status": "authoritative" }]
}
Get started

Give your agents a memory in minutes.