RAG vs Fine-Tuning vs Long Context

# Long Context with Anthropic Claude
import anthropic

client = anthropic.Anthropic()

# 1. Load your entire knowledge base into context
def load_documents(directory: str) -> str:
    """Load all documents into a single context string."""
    import os
    texts = []
    for filename in sorted(os.listdir(directory)):
        with open(os.path.join(directory, filename)) as f:
            texts.append(f"## {filename}\n{f.read()}")
    return "\n\n---\n\n".join(texts)

corpus = load_documents("./knowledge_base")
print(f"Corpus size: {len(corpus):,} characters")

# 2. Query with full context (simple!)
def long_context_query(question: str) -> str:
    response = client.messages.create(
        model="claude-opus-4-6-20260321",
        max_tokens=4096,
        messages=[
            {
                "role": "user",
                "content": f"""Here is a complete knowledge base:

<documents>
{corpus}
</documents>

Based on the documents above, answer this question:
{question}

Cite specific documents by name. If the answer spans multiple
documents, synthesize the information."""
            }
        ]
    )
    return response.content[0].text

# 3. Use prompt caching to amortize cost across queries
def cached_query(question: str) -> str:
    """Use prompt caching - corpus is cached after first call."""
    response = client.messages.create(
        model="claude-opus-4-6-20260321",
        max_tokens=4096,
        messages=[
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": f"<documents>\n{corpus}\n</documents>",
                        "cache_control": {"type": "ephemeral"}
                    },
                    {
                        "type": "text",
                        "text": f"Answer: {question}"
                    }
                ]
            }
        ]
    )
    return response.content[0].text