HN Explorer

Show HN: AI·rete·RAG – a Rete rule engine decides, RAG explains why (ai-rete-rag.com)

2 points by ZaharaHussain ↗ HN
Hi HN, I built ai·rete·rag because I kept seeing teams put an LLM in charge of decisions that need to be auditable (lending, fraud, clinical triage), then bolt on "guardrails" after the fact.

It runs the two in series instead:

1. A pure-Python Rete engine evaluates YAML rules against your facts. The verdict comes only from here. Same facts, same verdict, every time, with salience-based conflict resolution. 2. RAG retrieves passages from your own policy documents, and an LLM writes a plain-English explanation of the decision that was already made, citing those passages. It can't change the verdict.

A few things that went further than I expected: - Rules are a graph, not flat lists: nested all/any/not, and rules can assert facts that other rules consume (forward chaining). The decision trace shows the causal chain. - Audit mode records every rule evaluated, including the ones that didn't fire, condition by condition, with a snapshot of the rule set for replay. - Rules can steer retrieval (a fired rule narrows which documents get searched), and retrieved text can be turned into facts for the engine. - Non-technical authors can build rules in a visual editor, or paste a policy document and get LLM-drafted rules with citations. Drafts are never saved without review. YAML is still there for engineers.

The landing page has a live demo with no signup (8 demo domains: loan, fraud, clinical, insurance, legal, ops, e-commerce, blockchain). There's also an MCP server, so Claude and other agents can call /decide as a tool: `uvx ai-rete-rag-mcp`.

To be upfront: it's a hosted product with a free tier. The MCP client is open source (MIT, github.com/zaharajabeen13-create/ai-rete-rag-mcp); the engine and platform are not open source right now.

I'd especially like to hear from anyone who has had to explain an automated decision to a regulator or an auditor: what did they actually ask for?

33 comments

[ 1.6 ms ] story [ 166 ms ] thread
chews
It's been a while since I used a Rete rule engine, back in my days as math nerd at a mortgage bank (ooof we survived the 08 meltdown by the skin of our teeth) we used an internal fork of Drools and paired it with a merkle tree datasctructure to store lending decisions.... I knew blockchain was gonna be a thing when the genesis block referenced the bailout... I kick myself daily cause should've bought more bitcoin.
nickphx
I don't understand. If you're already building the rules, why have a clanker summarize and guess why the result happened when you can display the rules that passed/failed?
ZaharaHussain
Fair question, and you can do exactly that( response_mode:"verdict_only") makes no LLM call
lunatuna
I've used rete based CEP tool in the past for utility operations event management. We could tag the rule used to the event or event aggregate, but it was on a single node basis and a basic aggregate. If there was more coverage in understanding it would be a big help to operators. Based on say a set of outage events, can it further understand if it weather event, equipment failure, scheduled maintenance, etc. Linking it to possible operating procedures based on a broader understanding of the state would be very useful.
ZaharaHussain
If you still work near this, I'd genuinely like to hear what the rule set looked like - outage classification is the case I'd build the next demo domain around.
lunatuna
It’s been a long time since I did the work and lost for details. The biggest issue we had was noise reduction when there was an upstream event. Could be a network issue and so meter reads were not expected to come in, so don’t send someone out to check the meter or do deeper analysis. Storm or other related outage, both diagnose and then quiet follow on events with a service ticket to acknowledge. Further we would build the history of events and if a meter or device wasn’t responding without the other broader events then create a service ticket to analyze or to send a technician out. This was all custom rules. Not sure how transferable it is as different utilities will have different ways of viewing and actioning. Interconnected are somewhat different. North America is also very different than say Germany or Brazil which I had a little bit of experience with. Wish I had more details.
jimmySixDOF
I was expecting to see some Jev based influence on RETE/CEP rule handling ... the Typesafe.ai team have a lot more in the pipeline this use case is just getting started
awfm9
This is a really interesting concept. Could be quite useful for agents. I worked on something similar, but more abstract; you managed to take it to the next level.
lfdo0870
For that there's Python: with 30 lines you can automate this task. If you want I can send you the script I use. DM me if interested.
keparlak
The distinction you’ve drawn (the decision is made by the deterministic engine; the LLM merely explains it) matches something I’ve seen in another field. I was testing the agent’s memory: the schema of a vehicle changes mid-run, and the agent must make a call appropriate to the new schema at the end. Results: - A 20-line ‘the latest definition wins’ rule: 40/40 - Providing the entire history to the model: 38/40 - Retrieval-based memory: Mem0 (open-source) 3/15, TF-IDF + recency 1/40

When presented with the correct information, the model used it without issue. What it couldn’t reliably do was decide which information was still valid.

I’m curious about versioning on your end. When a rule changes, what happens to decisions made under the old rule? If someone asks six months later, “Why was this application rejected?”, does the explanation refer to the rule version at the time of the decision, or to the current one? Does the RAG side know that a policy passage has become invalid, or could it retrieve the old document to explain a new decision?

ZaharaHussain
Great data point, and I think it's the same failure mode: retrieval is good at "what's relevent", bad at "what's still in force". validity is a property you have to assign, not something similarility can recover.