This is a tiny inference engine (~120 lines of purely functional R5RS Scheme) I wrote a while back, when I was refreshing myself on how a Prolog interpreter works. I found several descriptions and examples of Prolog interpreters online, but none were quite what I wanted.

Cardboard Prolog lacks all the amenties the Prolog language proper, and it uses Scheme literals instead of Prolog syntax, but it does do the thing that's at the core of Prolog execution: deduction based on Horn clauses.

There are no comments, but there is a suite of tests. You can run the tests with (for example) Chicken Scheme, by running

csi -b test-cardboard-prolog.scm

I'll also try to briefly describe what's going on here.

A term in Cardboard Prolog is represented by a Scheme list, where atoms are symbols and variables are vectors of length 1 or 2. The first entry of a variable vector is a symbol giving the variable name, and the second entry is an index which is used to disambiguate different instances of a variable.

ground? and variable? are predicates on terms.

rename-term takes a term and returns a new term which is the same as the input term except that all variables are given new indexes. The purpose is to obtain a "fresh" version of the term with no bound variables.

collect-vars takes a term and returns a list of all variables found in it.

match-var and unify are mutually-recursive functions which implement unification. unify takes two patterns (terms which may contain variables) and returns a list of bindings if the each pattern matches the other, or #f if they cannot be matched. Such a list of bindings is called an "environment" (abbreviated env) in this code. Each binding is a two-element list of a variable, and the subterm that it matched with, which may itself be a variable, or contain variables.

Note that, for simplicitly, the unification algorithm here does not perform an occurs check. For the sake of correctness, it should perform one, but since it's very easy to implement and doesn't really add explanatory value to the exposition, I left it out. You can undertake adding one as an exercise, if you like.

expand takes a term and an environment and returns a new term which is the same as the input term except that all variables are replaced with the terms that they are bound to in the environment. subst is a helper function used by expand.

During the search process, a variable like #(X) will be instantiated to a variable like #(X 2) (where 2 indicates the depth of the search), and it is #(X 2) that will match a term, but this information is usually irrelevant to the user, for whom the report that #(X) matched would be more meaningful. collapse-env (with its helper functions expand-env and expand-binding) and restrict-to-vars are used clean up the output of the engine, and make its results more presentable to the user in this way.

search implements the core inference process. It is given a database (a list of facts and rules, where a fact is simply a rule with no premises), and a list of goals. It keeps track of the current environment (list of bindings) and the current search depth.

search tries to unify each rule in the database with the first goal of the current list of goals, under the current environment. If this succeeds, it takes the unifying environment (which we now call a "unifier"), expands the consequent of the rule and the remaining goals using the unifier, joins these together to obtain a new list of goals, and recursively calls itself with the new list of goals and the new environment, to continue to the search. If there are no more goals in the list to satisfy, the search was a success and the final unifying environment is returned.

But note that search might actually return to itself, because it calls itself recursively. So it returns a list of unifying environments, and collects these lists to ultimately return all of the successful searches in the database.

A real Prolog interpreter would do this piecemeal, asking the user if they want it to search for the next answer after each answer is found. For simplicitly, Cardboard Prolog always returns all the answers, and if there are infinitely many answers, this will simply not terminate.

(This design choice was for simplicitly, but it would certainly be an interesting exercise to rewrite it to work in the fashion of Prolog. Many of the descriptions I found online did describe how Prolog interpreters accomplish this, but none of them phrased it in terms of continuations, which is probably how you'd want to do it in Scheme.)

Finally, match-all is a driver function for search, and the main interface to the inference engine. It takes a database and a list of goals, and returns a list of comprehensible answers.