AI·rete·RAG splits decisions and explanations between a Rete engine and an LLM
AI·rete·RAG pairs a Python Rete rule engine that makes auditable decisions with a RAG and LLM layer that only explains them.
The developer wrote that the tool was built after seeing teams "put an LLM in charge of decisions that need to be auditable ... then bolt on guardrails after the fact." AI·rete·RAG instead runs two components in series. A pure-Python Rete engine evaluates YAML rules against a set of facts, and the verdict comes only from that engine. According to the post, the same facts produce the same verdict every time, with salience-based conflict resolution handling competing rules.
RAG sits downstream of the verdict. It retrieves passages from the user's own policy documents, and an LLM writes a plain-English explanation of the decision that has already been made, citing those passages. The language model cannot change the outcome.
The developer described several capabilities that went further than the original design. Rules are held as a graph rather than a flat list, supporting nested all, any and not operators, and a rule can assert facts that other rules then consume through forward chaining; the decision trace shows the causal chain. An audit mode records every rule that was evaluated, including the ones that did not fire, condition by condition, together with a snapshot of the rule set for replay.
Rules and retrieval interact in both directions. A fired rule can narrow which documents are searched, and retrieved text can be converted into facts for the engine.
Non-technical authors can build rules in a visual editor, or paste a policy document and receive LLM-drafted rules with citations. Drafts are not saved without review, and YAML remains available for engineers.
A live demo runs on the landing page without signup and covers eight domains: loan, fraud, clinical, insurance, legal, ops, e-commerce and blockchain. An MCP server, invoked with uvx ai-rete-rag-mcp, lets Claude and other agents call the /decide endpoint as a tool.
The developer said AI·rete·RAG is a hosted product with a free tier. The MCP client is open source under the MIT license at github.com/zaharajabeen13-create/ai-rete-rag-mcp, while the engine and the platform are not open source at present.
The post asked for feedback from people who have had to explain an automated decision to a regulator or an auditor, specifically what those reviewers asked for. The submission had drawn 5 points and no comments at the time of writing.