Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

openbkn-ai/bkn-foundry

AI Engineering

A backend foundation that models business semantics as ontologies and serves them to AI agents as governed context and executable actions.

1,399
139
Go
NOASSERTION
5 months old

What it does

BKN Foundry is the back end of OpenBKN, an ontology-driven business knowledge network. Business concepts are modeled as Markdown-based BKN definitions covering Data, Logic, Risk, and Action; the backend then turns those definitions into runtime services for agents. The project targets teams connecting proprietary systems and documents to LLM agents under enterprise governance, and it is backend-only: all interaction happens through the CLI, SDK, or API, with no bundled web console.

The engine is split into several parts: a Context Loader that recalls and ranks candidate knowledge for each reasoning turn, VEGA for data virtualization across heterogeneous sources, an Exec Factory for tools, MCP, and skills, BKN Safe for identity, permissions, and policy checks down to object and action level, and BKN Trace for tracing an action back to its knowledge node, data source, and tool. See the reference architecture for how these fit together.

The README reports benchmark results on 145 HR-scenario samples, including 99.31% accuracy against Dify and RAGFlow baselines, plus ablation tables showing schema preloading and a curated toolset each improving accuracy while reducing steps and tokens. These are author-reported measurements on a specific dataset with DeepSeek V3.2 and BGE M3-Embedding; they are not an independent evaluation.

Quick start

A full deployment runs on Linux through a Kubernetes stack managed by scripts in the separate openbkn-deploy repository. On the target host, clone that repository, run the preflight check, then the install script:

git clone https://github.com/openbkn-ai/openbkn-deploy.git
cd openbkn-deploy
chmod +x preflight.sh deploy.sh onboard.sh

sudo bash ./preflight.sh          # check only; exit 0 OK, 1 FAIL, 2 WARN
./deploy.sh openbkn install       # or pass --access_address and --api_server_address

The preflight script checks kernel, sysctl, containerd, kubectl, helm, Node, and BKN CLIs, and can apply fixes with --fix. The install script accepts --access_address for the address clients use and --api_server_address for a real NIC IP used by the Kubernetes API server. Verify with kubectl get nodes, kubectl get pods -A, and ./deploy.sh openbkn status.

After installing, run the bootstrap script on the same host to register an LLM and an embedding model; on a full install it also creates a test business user and assigns roles:

sudo bash ./onboard.sh

Then install the client CLI from bkn-sdk and log in:

npm install -g @openbkn/bkn-sdk
openbkn auth login https://<node-ip> -u test -p '<password>' -k
openbkn bkn list

-k skips TLS verification for self-signed certificates on a fresh deployment. onboard.sh needs sudo because it reads the config written by sudo deploy.sh under /root/.openbkn-ai/; re-running it is safe. Runnable end-to-end walkthroughs live in bkn-samples.

Using the SDK and CLI

The openbkn CLI is the main entry point: openbkn bkn list lists knowledge networks, openbkn bkn search <kn-id> "query" performs semantic search, and openbkn admin covers users, organizations, roles, models, and audit logs. A TypeScript client is available for programmatic access:

import { createClient } from "@openbkn/bkn-sdk";
const bkn = createClient({ baseUrl: "https://your-openbkn.com", token: process.env.BKN_TOKEN });
const networks = await bkn.kn.list({ limit: 10 });

For SSH, CI, or container environments, the CLI supports password sign-in via -u/-p, exporting and replaying credentials from a machine with a browser, and --no-browser with a pasted callback URL or code. An openbkn skill can be added to compatible agents with npx skills add https://github.com/openbkn-ai/bkn-sdk --skill openbkn, after which the agent can operate the platform on the user's behalf.

Constraints

  • Full installs target Linux; macOS is supported only for local development via kind.
  • Installation requires a Kubernetes cluster, containerd, kubectl, helm, and Node, as checked by preflight.sh.
  • Authentication and authorization are always on; built-in accounts (system, admin, security, audit) enforce separation of duties, and operators should use individual accounts.
  • New users receive a generated initial password (returned once) and must change it at first sign-in.
  • The repository is multi-licensed per component: Apache 2.0 applies to some files, while OpenBKN net-new files use a modified Apache 2.0 with additional conditions. The NOTICE file is the authoritative breakdown.

HotRepo's take

On the Go daily list for 2026-11-16, the project recorded 1,398 stars (up 48 that day), 138 forks, and a best position of 9. That is a single snapshot, so it establishes presence on the daily Go list on that date, not a trend. The project's own framing is worth noting: 99%+ accuracy alongside higher token usage and latency than the BiSheng baseline in its own table, which suggests the architecture trades inference cost for retrieval quality. For teams evaluating ontology backends for agents, the benchmarks and ablation numbers are a starting point for reproducing on their own data rather than settled proof.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. NVIDIA/auto-ontologyBuilds a certified semantic ontology over enterprise databases so LLM agents can generate validated SQL.Python12,983
  2. gogoymh/auto-ontology-mcpAn MCP server that builds a semantic layer over databases and a notes knowledge base, answering questions with validated SQL.Python105