Open-source vectors with persistent context

import chromadb
import httpx
import os

# Chroma: Simple local vector storage
chroma_client = chromadb.Client()  # or PersistentClient for disk storage
collection = chroma_client.get_or_create_collection(
    name="project_docs",
    metadata={"hnsw:space": "cosine"}
)

# Stompy: Simple cloud memory
STOMPY_TOKEN = os.environ["STOMPY_TOKEN"]

async def stompy_recall(topic: str) -> str:
    async with httpx.AsyncClient() as client:
        response = await client.post(
            "https://mcp.stompy.ai/sse",
            headers={"Authorization": f"Bearer {STOMPY_TOKEN}"},
            json={"tool": "recall_context", "topic": topic}
        )
        data = response.json()
        return data.get("content", "")

async def contextual_rag(user_query: str):
    # Get relevant documents from Chroma
    results = collection.query(
        query_texts=[user_query],
        n_results=5,
        include=["documents", "metadatas", "distances"]
    )

    # Get project context from Stompy
    project_context = await stompy_recall("architecture_decisions")
    tech_constraints = await stompy_recall("tech_stack")

    # Build informed prompt
    retrieved_docs = "\n---\n".join(results["documents"][0])

    prompt = f"""You are helping with a project. Here's the context:

PROJECT DECISIONS:
{project_context}

TECHNICAL CONSTRAINTS:
{tech_constraints}

RETRIEVED DOCUMENTATION:
{retrieved_docs}

USER QUESTION: {user_query}

Answer considering our specific project context:"""

    return await call_llm(prompt)

async def save_rag_insight(insight_type: str, content: str):
    """Save development insights for future sessions."""
    async with httpx.AsyncClient() as client:
        await client.post(
            "https://mcp.stompy.ai/sse",
            headers={"Authorization": f"Bearer {STOMPY_TOKEN}"},
            json={
                "tool": "lock_context",
                "topic": f"rag_development_{insight_type}",
                "content": content,
                "tags": "rag,development,chroma"
            }
        )

# Examples of what to save:
await save_rag_insight("chunking", """
Chunking Strategy Decision:
- 512 tokens with 50 token overlap works best for our docs
- Larger chunks lost specificity for code questions
- Sentence-based splitting broke code blocks badly
""")

await save_rag_insight("retrieval", """
Retrieval Tuning:
- n_results=5 is sweet spot (3 too few, 10 adds noise)
- Metadata filter on 'doc_type' helps for API questions
- Distance threshold 0.7 filters low-quality matches
""")