python
from sentence_transformers import CrossEncoder

# Step 1: Hybrid retrieval
bm25_results = bm25_search(query, top_k=20)
vector_results = vector_search(query, top_k=20)

# Step 2: Merge and deduplicate
candidates = merge_results(bm25_results, vector_results)

# Step 3: Cross-encoder reranking
reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-12-v2")
scores = reranker.predict([(query, doc.text) for doc in candidates])

# Step 4: Take top results
reranked = sorted(zip(candidates, scores), key=lambda x: x[1], reverse=True)
final_context = [doc for doc, _ in reranked[:5]]

python
# 1. Extract entities and relationships using LLM
entities = extract_entities(document)
# Output: [("OpenAI", "COMPANY"), ("GPT-4", "MODEL"), ...]

# 2. Build knowledge graph
graph.add_edge("OpenAI", "GPT-4", relation="CREATED")
graph.add_edge("GPT-4", "Transformer", relation="BASED_ON")

# 3. Community detection for summarization
communities = detect_communities(graph)

# 4. Query using graph traversal + vector search
results = graph_rag_query(
    "What models has OpenAI created?",
    graph=graph,
    vector_store=vector_store
)

python
class AgenticRAG:
    def answer(self, question: str) -> str:
        plan = self.llm.plan_retrieval(question)
        
        context = []
        for step in plan:
            if step.type == "SEARCH":
                results = self.search(step.query)
                context.extend(results)
            elif step.type == "FILTER":
                context = self.filter(context, step.criteria)
            elif step.type == "SYNTHESIZE":
                partial = self.llm.synthesize(context)
                context = [partial]
        
        return self.llm.generate_final(question, context)

python
# Add document-level context to each chunk
def contextual_chunk(document: str, chunks: list) -> list:
    doc_summary = llm.summarize(document)
    
    enhanced_chunks = []
    for chunk in chunks:
        enhanced = f"""
        Document Context: {doc_summary}
        Section: {chunk.section_title}
        Content: {chunk.text}
        """
        enhanced_chunks.append(enhanced)
    
    return enhanced_chunks