In this tutorial, we are walking through a hands-on fusion of symbolic logic and generative AI. We set up PySwip to embed a Prolog knowledge base, wrap its predicates as LangChain tools, and then wire everything into a ReAct-style agent. Along the way, we are crafting family-relationship rules, mathematical predicates like factorial, and list utilities, then letting the agent plan, call tools, and reason over the results. By the end of the setup, we can issue natural-language questions and watch the agent translate them into precise Prolog queries, stitch together multi-step answers, and return structured JSON-backed insights.
!apt-get install swi-prolog -y!pip install pyswip langchain-google-genai langgraph langchain-coreWe install SWI-Prolog with apt-get and then add pyswip, LangChain’s Google GenAI wrapper, LangGraph, and core LangChain packages via pip so we can bridge Prolog logic with our Gemini-powered agent. With these dependencies in place, we’re ready to code, query, and orchestrate reasoning end to end.
import osfrom pyswip import Prologfrom langchain_google_genai import ChatGoogleGenerativeAIfrom langchain_core.messages import HumanMessagefrom langchain_core.tools import toolfrom langgraph.prebuilt import create_react_agentimport jsonGOOGLE_API_KEY = "Use Your Own API Key Here"os.environ["GOOGLE_API_KEY"] = GOOGLE_API_KEYllm = ChatGoogleGenerativeAI(model="gemini-1.5-flash", temperature=0)We load our core stack, including PySwip for Prolog, LangChain and LangGraph for tooling, and Gemini 1.5 Flash for LLM power. We then set the GOOGLE_API_KEY environment variable so the model can authenticate. With the LLM initialized at zero temperature, we’re primed to get deterministic, logic-grounded answers from our agent.
class AdvancedPrologInterface: def __init__(self): self.prolog = Prolog() self._load_knowledge_base() def _load_knowledge_base(self): """Load comprehensive Prolog knowledge base""" rules = [ "parent(john, mary, alice)", "parent(john, mary, bob)", "parent(bob, susan, charlie)", "parent(alice, david, emma)", "parent(charlie, lisa, frank)", "male(john)", "male(bob)", "male(david)", "male(charlie)", "male(frank)", "female(mary)", "female(alice)", "female(susan)", "female(emma)", "female(lisa)", "grandparent(X, Z) :- parent(X, _, Y), parent(Y, _, Z)", "sibling(X, Y) :- parent(P1, P2, X), parent(P1, P2, Y), X \\= Y", "uncle(X, Y) :- sibling(X, Z), parent(Z, _, Y), male(X)", "aunt(X, Y) :- sibling(X, Z), parent(Z, _, Y), female(X)", "cousin(X, Y) :- parent(P1, _, X), parent(P2, _, Y), sibling(P1, P2)", "factorial(0, 1)", "factorial(N, F) :- N > 0, N1 is N - 1, factorial(N1, F1), F is N * F1", "list_member(X, [X|_])", "list_member(X, [_|T]) :- list_member(X, T)", "list_length([], 0)", "list_length([_|T], N) :- list_length(T, N1), N is N1 + 1", "animal(dog)", "animal(cat)", "animal(whale)", "animal(eagle)", "mammal(dog)", "mammal(cat)", "mammal(whale)", "bird(eagle)", "bird(sparrow)", "can_fly(eagle)", "can_fly(sparrow)", "can_swim(whale)", "can_swim(fish)", "aquatic_mammal(X) :- mammal(X), can_swim(X)" ] for rule in rules: try: self.prolog.assertz(rule) except Exception as e: print(f"Warning: Could not assert rule '{rule}': {e}") def query(self, query_string): """Execute Prolog query and return results""" try: results = list(self.prolog.query(query_string)) return results if results else [{"result": "No solutions found"}] except Exception as e: return [{"error": f"Query failed: {str(e)}"}]We wrap SWI-Prolog in an AdvancedPrologInterface, load a rich rule/fact base on init, and assert each clause safely. We then expose a query() method that runs any Prolog goal and returns JSON-friendly results (or a clear error/no-solution message), allowing us to drive logic queries directly from Python.
prolog_interface = AdvancedPrologInterface()@tooldef family_relationships(query: str) -> str: """ Query family relationships in Prolog format. Examples: 'parent(john, mary, X)', 'sibling(X, Y)', 'grandparent(X, charlie)' """ results = prolog_interface.query(query) return json.dumps(results, indent=2)@tooldef mathematical_operations(operation: str, number: int) -> str: """ Perform mathematical operations using Prolog. Supported operations: 'factorial' Example: operation='factorial', number=5 """ if operation == "factorial": query = f"factorial({number}, Result)" results = prolog_interface.query(query) return json.dumps(results, indent=2) else: return json.dumps([{"error": f"Operation '{operation}' not supported"}])@tooldef advanced_queries(query_type: str, entity: str = "") -> str: """ Perform advanced relationship queries. Types: 'all_children', 'all_grandchildren', 'all_siblings', 'all_cousins' """ queries = { 'all_children': f"parent(_, _, {entity})" if entity else "parent(_, _, X)", 'all_grandchildren': f"grandparent(_, {entity})" if entity else "grandparent(_, X)", 'all_siblings': f"sibling({entity}, X)" if entity else "sibling(X, Y)", 'all_cousins': f"cousin({entity}, X)" if entity else "cousin(X, Y)" } if query_type in queries: results = prolog_interface.query(queries[query_type]) return json.dumps(results, indent=2) else: return json.dumps([{"error": f"Query type '{query_type}' not supported"}])We instantiate AdvancedPrologInterface and then wrap its queries as LangChain tools, such as family_relationships, mathematical_operations, and advanced_queries, so that we can call precise Prolog goals from natural language. We define each tool to format and dispatch the right query (such as factorial/2 or cousin lookups) and return clean JSON, allowing our agent to orchestrate logic calls seamlessly.
tools = [family_relationships, mathematical_operations, advanced_queries]agent = create_react_agent(llm, tools)def run_family_analysis(): """Comprehensive family relationship analysis""" print("
Family Relationship Analysis") print("=" * 50) queries = [ "Who are all the parents in the family database?", "Find all grandparent-grandchild relationships", "Show me all the siblings in the family", "Who are John and Mary's children?", "Calculate the factorial of 6 using Prolog" ] for i, query in enumerate(queries, 1): print(f"\n
Query {i}: {query}") print("-" * 30) try: response = agent.invoke({"messages": [("human", query)]}) answer = response["messages"][-1].content print(f"
Response: {answer}") except Exception as e: print(f"
Error: {str(e)}")def demonstrate_complex_reasoning(): """Show advanced multi-step reasoning""" print("\n
Complex Multi-Step Reasoning") print("=" * 40) complex_query = """ I want a complete family tree analysis. Please: 1. List all parent-child relationships 2. Identify all grandparent relationships 3. Find any uncle/aunt relationships 4. Show cousin relationships 5. Calculate factorial of 4 as a bonus math operation """ print(f"Complex Query: {complex_query}") print("-" * 40) try: response = agent.invoke({"messages": [("human", complex_query)]}) print(f"
Comprehensive Analysis:\n{response['messages'][-1].content}") except Exception as e: print(f" Error in complex reasoning: {str(e)}")def interactive_prolog_session(): """Interactive Prolog knowledge base exploration""" print("\n
Interactive Prolog Explorer") print("Ask about family relationships, math operations, or general queries!") print("Type 'examples' to see sample queries, 'quit' to exit") print("-" * 50) examples = [ "Who are Bob's children?", "Find all grandparents in the family", "Calculate factorial of 5", "Show me all cousin relationships", "Who are Alice's siblings?" ] while True: user_input = input("\n
You: ") if user_input.lower() == 'quit': print("
Goodbye!") break elif user_input.lower() == 'examples': print("
Example queries:") for ex in examples: print(f" • {ex}") continue try: response = agent.invoke({"messages": [("human", user_input)]}) print(f" AI: {response['messages'][-1].content}") except Exception as e: print(f" Error: {str(e)}")We register our three Prolog tools, spin up a ReAct agent around Gemini, and then script helper routines, run_family_analysis, demonstrate_complex_reasoning, and an interactive loop, to fire natural-language queries that the agent translates into Prolog calls. This way, we test simple prompts, multi-step reasoning, and live Q&A, all while keeping the logic layer transparent and debuggable.
def test_direct_queries(): """Test direct Prolog queries for verification""" print("\n
Direct Prolog Query Testing") print("=" * 35) test_queries = [ ("parent(john, mary, X)", "Find John and Mary's children"), ("grandparent(X, charlie)", "Find Charlie's grandparents"), ("sibling(alice, X)", "Find Alice's siblings"), ("factorial(4, X)", "Calculate 4 factorial"), ("cousin(X, Y)", "Find all cousin pairs") ] for query, description in test_queries: print(f"\n {description}") print(f"Query: {query}") results = prolog_interface.query(query) print(f"Results: {json.dumps(results, indent=2)}")def main(): """Main demonstration runner""" if GOOGLE_API_KEY == "YOUR_GEMINI_API_KEY_HERE": print("
Please set your Gemini API key in Cell 3!") print("Get it from: https://aistudio.google.com/app/apikey") return print("
Advanced Prolog + Gemini Integration") print("Using PySwip for stable Prolog integration") print("=" * 55) test_direct_queries() run_family_analysis() demonstrate_complex_reasoning() def show_mathematical_capabilities(): """Demonstrate mathematical reasoning with Prolog""" print("\n
Mathematical Reasoning with Prolog") print("=" * 40) math_queries = [ "Calculate factorial of 3, 4, and 5", "What is the factorial of 7?", "Show me how factorial calculation works step by step" ] for query in math_queries: print(f"\n
Math Query: {query}") try: response = agent.invoke({"messages": [("human", query)]}) print(f"
Result: {response['messages'][-1].content}") except Exception as e: print(f" Error: {str(e)}")if __name__ == "__main__": main() show_mathematical_capabilities() print("\n
Tutorial completed successfully!") print("
Key achievements:") print(" • Integrated PySwip with Gemini AI") print(" • Created advanced Prolog reasoning tools") print(" • Demonstrated complex family relationship queries") print(" • Implemented mathematical operations in Prolog") print(" • Built interactive AI agent with logical reasoning") print("\n Try extending with your own Prolog rules and facts!")We wire everything together in main() to verify our Prolog goals, run the family analysis, and showcase multi-step reasoning, then show_mathematical_capabilities() stresses factorial queries from natural language. We conclude by printing a quick recap of what we’ve built so far, allowing us to confidently extend the stack with new rules or swap models next.
In conclusion, we have demonstrated that symbolic reasoning and LLMs complement each other beautifully: Prolog guarantees correctness on well-defined logic, while Gemini handles flexible language understanding and orchestration. We are leaving with a working scaffold, direct Prolog queries for verification, tool-wrapped predicates for agents, and demo functions for complex family tree and mathematical analyses. From here, we are ready to expand the knowledge base, add new domains (such as finance rules, game logic, and knowledge graphs), or swap in different LLMs. We are also positioned to expose this stack via an interactive UI or API, allowing others to explore logic-guided AI in real-time.
Check out the Full Codes. All credit for this research goes to the researchers of this project.
Meet the AI Dev Newsletter read by 40k+ Devs and Researchers from NVIDIA, OpenAI, DeepMind, Meta, Microsoft, JP Morgan Chase, Amgen, Aflac, Wells Fargo and 100s more [SUBSCRIBE NOW]
The post A Coding Guide to Build a Tool-Calling ReAct Agent Fusing Prolog Logic with Gemini and LangGraph appeared first on MarkTechPost.
