Real World Example

Bank and fintech chatbots (HDFC EVA, ICICI iPal, PhonePe and Paytm help) combine everything in this module: per-session memory, streaming answers, strict error handling, and a hand-over to a human. The ShopEase chat architecture:

Next.js chat UI ──POST /api/chat (SSE)──► FastAPI
                                           │  load history + summary (Redis, key chat:c-77, TTL 24h)
                                           │  retrieve policy chunks (RAG)  +  tools (orders)
                                           ▼
                                     Claude API (stream)
                                           │
                     events ◄──────────────┘   save history; summarize when > 20 turns
                     "Talk to a human" button → creates a Freshdesk/Zendesk ticket with the summary

import json
import redis.asyncio as redis

r = redis.from_url("redis://localhost:6379")
TTL = 24 * 3600

async def load_history(cid: str) -> list:
    raw = await r.get(f"chat:{cid}")
    return json.loads(raw) if raw else []

async def save_history(cid: str, history: list) -> None:
    serializable = [
        {"role": m["role"],
         "content": m["content"] if isinstance(m["content"], str)
                    else [b.model_dump() if hasattr(b, "model_dump") else b for b in m["content"]]}
        for m in history
    ]
    await r.set(f"chat:{cid}", json.dumps(serializable), ex=TTL)