✨ add ICV RAG example

Author: HamedBabaei

examples/icv_rag_matching.py

@@ -0,0 +1,93 @@
+import json
+from ontoaligner.ontology import MaterialInformationMatOntoOMDataset
+from ontoaligner.utils import metrics, xmlify
+from ontoaligner.ontology_matchers import FalconLLMBERTRetrieverICVRAG
+from ontoaligner.encoder import ConceptRAGEncoder
+from ontoaligner.postprocess import rag_hybrid_postprocessor
+
+# Step 1: Initialize the Ontology Matching Task
+# The MaterialInformationMatOntoOMDataset object is created to start the ontology matching task
+task = MaterialInformationMatOntoOMDataset()
+print("Test Task:", task)
+
+# Step 2: Collect the dataset for ontology matching
+# This collects the source ontology, target ontology, and reference alignments required for the matching process
+dataset = task.collect(
+    source_ontology_path="../assets/MI-MatOnto/mi_ontology.xml",  # Path to the source ontology file
+    target_ontology_path="../assets/MI-MatOnto/matonto_ontology.xml",  # Path to the target ontology file
+    reference_matching_path="../assets/MI-MatOnto/matchings.xml"  # Path to the reference alignments
+)
+
+# Step 3: Initialize the ConceptRAGEncoder model
+# This encoder will process the source and target ontologies for the matching process
+encoder_model = ConceptRAGEncoder()
+
+# Step 4: Encode the ontologies
+# This encodes the source, target, and reference ontologies into a format that can be used by the model
+encoded_ontology = encoder_model(source=dataset['source'], target=dataset['target'], reference=dataset['reference'])
+
+# Step 5: Define model configuration for FalconLLMBERTRetrieverICVRAG
+# This includes the device settings, retrieval configuration, and LLM configuration for the matching model
+config = {
+    "retriever_config": {
+        "device": 'cuda',  # Specify the device for computation
+        "top_k": 5,  # Number of top candidates to retrieve
+        "threshold": 0.1,  # Threshold for IR scores
+        # "openai_key": ""  # Uncomment to use OpenAI models
+    },
+    "llm_config": {
+        "device": "cuda",  # Specify the device for computation
+        "max_length": 300,  # Max length for LLM input
+        "max_new_tokens": 10,  # Max new tokens for generation
+        "huggingface_access_token": "",  # Huggingface access token for restricted models
+        "device_map": 'balanced',  # Device mapping strategy
+        "batch_size": 32,  # Batch size for inference
+        "answer_set": {
+            "yes": ["yes", "correct", "true", "positive", "valid"],
+            "no": ["no", "incorrect", "false", "negative", "invalid"]
+        }
+        # "openai_key": ""  # Uncomment to use OpenAI models
+    }
+}
+
+# Step 6: Initialize the Matching Model (FalconLLMBERTRetrieverICVRAG)
+# This step creates an instance of the FalconLLMBERTRetrieverICVRAG model with the configuration provided
+model = FalconLLMBERTRetrieverICVRAG(**config)
+
+# Step 7: Load pre-trained models (LLM and IR models)
+# Loads the pre-trained Falcon LLM and information retrieval models
+model.load(
+    llm_path="tiiuae/falcon-7b",  # Path to the pre-trained LLM
+    ir_path="all-MiniLM-L6-v2"  # Path to the IR model
+)
+
+# Step 8: Generate predictions using the model
+# This generates predictions for the ontology matching task using the encoded data
+predicts = model.generate(input_data=encoded_ontology)
+
+# Step 9: Post-process the predictions using a hybrid approach
+# This applies a post-processing step to refine the predictions based on IR and LLM confidence thresholds
+hybrid_matchings, hybrid_configs = rag_hybrid_postprocessor(
+    predicts=predicts,
+    ir_score_threshold=0.4,  # IR score threshold for filtering
+    llm_confidence_th=0.5  # LLM confidence threshold for filtering
+)
+
+# Step 10: Evaluate the hybrid matchings
+# This generates an evaluation report comparing the predicted matchings to the reference matchings
+evaluation = metrics.evaluation_report(predicts=hybrid_matchings, references=dataset['reference'])
+print("Hybrid Matching Evaluation Report:", json.dumps(evaluation, indent=4))
+
+# Step 11: Print the hybrid matching configuration
+# This outputs the configuration used for hybrid matching
+print("Hybrid Matching Obtained Configuration:", hybrid_configs)
+
+# Step 12: Convert the hybrid matchings to XML format
+# This generates an XML representation of the final ontology matchings
+xml_str = xmlify.xml_alignment_generator(matchings=hybrid_matchings)
+
+# Step 13: Save the XML output to a file
+# The resulting XML string is saved to an XML file for further use
+output_file_path = "matchings.xml"
+with open(output_file_path, "w", encoding="utf-8") as xml_file:
+    xml_file.write(xml_str)