markdownmarty.com

What It Is

Markdown Marty defines the Markdown files themselves as the “model.” It is not a traditional large language model, nor does it perform RAG-style vector retrieval. At inference time, it explicitly makes 0 LLM calls and uses no neural networks, GPUs, embeddings, or external APIs. Instead, it reads Markdown nodes from disk, generates answers through pattern matching and template substitution, and returns cited sources.

Core Capabilities and Output Quality

Its main selling points are auditability and determinism. The text states that the runtime is about 700 lines of Python, with cold-cache queries taking around 30ms and hot-cache queries around 2ms, and that it can run Bible-scale corpora on a laptop. Every answer includes a provenance file path, and cited files must actually exist; if a corpus file is deleted, the behavior must change. This design can significantly reduce the “hallucination” problem common in traditional LLMs, but only if the knowledge has already been structured into Markdown and the templates and node organization are well designed.

API, Privacy, and Architectural Constraints

Markdown Marty provides an HTTP API and requires every response to include both the answer and the referenced corpus files. It also constrains the architecture with six tests: no LLM imports, no embedding/vector calls, no subprocess usage, verification that cited files exist, verification that the corpus affects the output, and a requirement that the API provide sources. This is friendly to compliance and auditing. Because inference does not call external APIs, it is suitable for local knowledge-base scenarios, but the text does not provide a formal privacy policy or security certification information.

Pricing and Ideal Users

The text indicates that the framework is open source and provides GitHub source code. It does not disclose the license, commercial pricing, hosted versions, or paid support. It is suitable for developers, knowledge engineering teams, and structured-knowledge scenarios where hallucinations are unacceptable, such as drug interactions, building codes, repair manuals, and case-law citations.

Pros, Cons, and Access from China

Its advantages are low cost, low latency, local deployment, and clear sourcing. Its drawbacks are that it is not a general-purpose AI system, cannot handle open-ended reasoning or creative generation, and its effectiveness depends heavily on how the corpus is modeled. Chinese-language support is not mentioned, and information on access from China or payment is also unknown. If you need a more general-purpose Chinese knowledge-base solution, you can compare it with Dify, RAGFlow, LangChain, LlamaIndex, and similar tools.