Knowledge Copilot

For enterprise teams& AI agents·Open source · Self-hosted

Context your agents and your team can defend.

The trust layer for enterprise knowledge. Agents ask, Knowledge Copilot grounds the answer in your documents, cites it, scores it, and keeps the context across the task. When the evidence isn’t there, it declines.

Run it in 5 minutes Star on GitHub
AgentsSupport AgentAnswers customer queriesResearch AgentFinds insights across docsOps AgentAssists internal workflowsKnowledge CopilotEnterprise Knowledge PlatformRetrieveFind knowledgeUnderstandExtract & reasonValidateCheck confidenceRespondCite sourcesAccess ControlCitationsConfidenceMemoryContinuous learning& knowledge base1 · Ask / Query3 · Store & Update2 · Context LoopFeedback improves future responses

0%

Grounded answer rate

0%

Citations hit the right document

0%

False answers on trap questions

0.0s

Median answer latency

Trust Bench v1 — a 35-question public benchmark with deliberate trap questions, versioned in bench/. Self-measured and reproducible against any deployment.

From raw documents to a trusted answer.

Every question runs the same governed pipeline. Nothing is answered that the evidence can’t support — and every step is inspectable.

Your documents
Structure-aware chunking
Embed + index
Hybrid retrieval + rerank
Ground, cite & score
Trusted answer
The trusted answer reaches your team, in the appand your agents, over the API & MCP

It answers — or it admits it can’t.

The same question, two honest outcomes. Grounded when the evidence is there; a clean refusal when it isn’t.

knowledge-copilot · live behavior

What monthly uptime does our SLA commit to?

Searching 5 documents…

Both behaviors are measured — see the numbers below.

EKCKnowledgeMemorySearchAgentsAPIsGovernanceENTERPRISE KNOWLEDGE COPILOT

The platform

One grounded core. Six ways in.

People ask in the app. Agents call the OpenAI-compatible endpoint or the MCP server. Everyone gets the same governed, cited knowledge.

  • KnowledgeDocuments, wikis, and drives become one cited corpus
  • MemoryScoped, expiring memory your agents can trust
  • SearchHybrid retrieval: dense + BM25 + cross-encoder rerank
  • AgentsOpenAI-compatible API and an MCP server built in
  • GovernanceConfidence gates, human review, full audit trail
  • APIsEverything in the UI is a documented REST endpoint

A trust layer, not just a chatbot.

The things enterprises actually need before they let AI touch their knowledge.

Confidence scoring

Every answer carries a composite score with retrieval, groundedness, and citation sub-scores.

Inline citations

Each claim links to the exact passage — click through to the highlighted source.

Honest refusals

When the evidence isn't there, it declines instead of guessing. No answer beats a wrong one.

Human review queue

Low-confidence answers route to a reviewer; every resolution is audited.

Agent memory & context

Scoped, expiring memory and token-budgeted context packs your agents can trust.

Reproducible evals

Built-in evaluation datasets and the open Trust Bench keep quality measurable.

Connect your sources once.

Click-to-connect OAuth for Drive and Notion, an API-token for Confluence, folders for everything else. Re-syncs are idempotent.

Google Drive Notion Confluence Server folder Slack SharePoint Dropbox Gmail Google Drive Notion Confluence Server folder Slack SharePoint Dropbox Gmail

For your agents

Point any agent at your knowledge.

One line changes the base URL. Your existing OpenAI-SDK code keeps working — but now answers are grounded in your documents and carry trust signals your workflow can branch on. There’s an MCP server too, for Claude and other MCP clients.

  • Grounded, cited completions at /v1/chat/completions
  • Trust signals ride in the ekc extension object
  • MCP tools: ask, search, remember, recall, write_knowledge

python

from openai import OpenAI

client = OpenAI(
    base_url="https://kb.your-company.com/v1",
    api_key="ekc_...",              # a key you mint
)
resp = client.chat.completions.create(
    model="legal",                  # a domain profile
    messages=[{"role": "user", "content": "..."}],
)
# resp.ekc -> { answered, confidence, citations, needs_review }
Self-hosted — data never leaves your server OIDC SSO + role-based access, free in OSS Full audit trail on every answer & review CPU-only — no GPU required

Yours in three lines.

CPU-only Docker Compose: Postgres, Redis, Qdrant, OCR, Office parsing, and the web app. No API keys required — local models by default, cloud LLMs one env var away.

bash

git clone https://github.com/malakazlan/Enterprise-Knowledge-Copilot.git
cd Enterprise-Knowledge-Copilot
docker compose -f infra/docker-compose.yml up -d

Open http://localhost:8000 — the first account registered becomes the administrator.

Give your knowledge a conscience.

Grounded answers, honest refusals, and a paper trail — for the people and the agents that depend on what your organization knows.

Get started Star on GitHub