ballast

Pull a quantized knowledge corpus and ground any local model. Works with Ollama and every OpenAI-compatible or MCP-capable client.

uvx --from git+https://github.com/OpenBallast/ballast-cli openballast pull --level 3
uvx --from git+https://github.com/OpenBallast/ballast-cli openballast serve

(PyPI package coming — it becomes just uvx openballast.)

• Ollama users: point your client's base URL at http://localhost:11435/v1 instead of http://localhost:11434/v1 — done. Every chat request is grounded with corpus facts before your model sees it. No tool calling needed, works with any model size.

• MCP users (Claude Desktop, LM Studio, Cline, Goose): add to your MCP config:

  { "ballast": { "command": "uvx",
                 "args": ["--from", "git+https://github.com/OpenBallast/ballast-cli",
                          "openballast", "mcp"] } }

• Smoke test:

  uvx --from git+https://github.com/OpenBallast/ballast-cli openballast lookup "Where was Douglas Adams born?"

What you're downloading

Ballast T0: 25.4M entities and 197M facts from Wikidata (CC0), quantized into nested levels — pick your knowledge size like you pick a GGUF quant:

Levels are nested: pull --level 5 after pull --level 3 downloads only the new buckets. Everything runs offline after the pull — no network at answer time.

Measured effect (details: thesis): a 2B model + ~180 MB of ballast exceeds a 12B model's factual accuracy; hallucination on factual probes drops ~3×.

Related MCP server: Open Research KB

Commands

ballast pull  --level 3      # download / upgrade the corpus
ballast serve                # OpenAI grounding proxy :11435 + MCP http :11436
ballast mcp                  # MCP on stdio (for client configs)
ballast lookup "question"    # print the evidence blocks for a question
ballast status               # installed levels and sizes

BALLAST_HOME overrides the storage location (default ~/.ballast).

How it works

serve intercepts POST /v1/chat/completions, mines entity mentions from your last message, resolves them against the local corpus (normalized label/alias match), and prepends the matching facts as a system message. Everything else — including streaming — passes through untouched. The MCP server exposes the same three tools (resolve, evidence, lookup) as the hosted demo endpoint (mcp.openballast.org).

Apache-2.0. Corpus data: CC0 (Wikidata contributors).