BetterDB × Krish Naik Academy

AI Infrastructure Observability — RAG Pipeline Demo

"Your AI App's Redis is Broken — And You Don't Know It"

A collaboration between BetterDB and Krish Naik

📺 Watch the Video · 📖 Full Step-by-Step Guide · 🌐 BetterDB

---

The Problem

You ship a RAG pipeline. It uses Redis for semantic caching, agent memory, and rate limiting. Three weeks later:

• A LangChain agent runs overnight — 10,000 HSET commands on one session key
• Your semantic cache grows to 450 MB — nobody set TTLs
• Someone hits your LLM rate limiter 10,000+ times in 30 seconds
• RAG pipeline p95 latency creeps up — HGETALL is the bottleneck, invisible in aggregate

You open CloudWatch. You see a spike. The Redis slowlog is already overwritten — 128 entries, gone in 0.1 seconds at 1,000 cmd/s.

BetterDB persists everything. Query it hours later, in plain English, from Claude Code.

---

What's in This Repo

A minimal FastAPI RAG application that generates real Redis keys so BetterDB has actual data to monitor — no fake seeding, no mock data.

betterdb-yt-collab/
├── rag/
│   ├── config.py       ← pydantic-settings + Redis + OpenAI clients
│   ├── pipeline.py     ← ingest, retrieve, semantic cache, rate limit, session
│   └── main.py         ← FastAPI: POST /ingest  POST /query  GET /stats  GET /health
├── docker-compose.yml  ← local Valkey (alternative to Upstash)
├── step-by-step.md     ← full demo walkthrough with all commands
├── .env.example        ← credential template
└── pyproject.toml      ← Python dependencies

---

Redis Key Patterns Generated

Key Pattern	Written by	TTL	BetterDB Feature
rag:doc:{sha256}	POST /ingest	None — the bug	Feature 2, 5
semantic_cache:{md5}	POST /query	None — the bug	Feature 2
rate_limit:user_{id}:minute	POST /query	60s ✓	Feature 4
rate_limit:user_{id}:hour	POST /query	3600s ✓	Feature 4
langchain:memory:session:{id}	POST /query	None — the bug	Feature 3

---

Quick Start

1. Install dependencies

uv venv --python 3.12 && source .venv/bin/activate
uv pip install fastapi uvicorn pypdf numpy pydantic-settings python-multipart openai redis python-dotenv

2. Configure .env

cp .env.example .env
# Fill in: OPENAI_API_KEY, REDIS_URL, BETTERDB_TOKEN

Option A — Upstash Redis (cloud, zero infra):

REDIS_URL=rediss://default:YOUR_TOKEN@quiet-flea-79665.upstash.io:6379

Option B — Local Valkey (Docker, better for slowlog demo):

REDIS_URL=redis://localhost:6379

docker-compose up -d   # start Valkey
docker-compose ps      # health check

3. Connect Redis to BetterDB cloud

BetterDB → Manage Connections → + Add Connection → Via Agent → copy token.

Upstash:

docker run -d --name betterdb-agent \
  -e VALKEY_HOST=quiet-flea-79665.upstash.io \
  -e VALKEY_PORT=6379 \
  -e VALKEY_TLS=true \
  -e VALKEY_USERNAME=default \
  -e VALKEY_PASSWORD=YOUR_UPSTASH_TOKEN \
  -e BETTERDB_CLOUD_URL=wss://betterdb-test1.app.betterdb.com/agent/ws \
  -e BETTERDB_TOKEN=YOUR_BETTERDB_AGENT_TOKEN \
  betterdb/agent:latest

Local Valkey:

docker run -d --name betterdb-agent-local \
  -e VALKEY_HOST=host.docker.internal \
  -e VALKEY_PORT=6379 \
  -e BETTERDB_CLOUD_URL=wss://betterdb-test1.app.betterdb.com/agent/ws \
  -e BETTERDB_TOKEN=YOUR_BETTERDB_AGENT_TOKEN \
  betterdb/agent:latest

4. Start FastAPI

uvicorn rag.main:app --reload --port 8000

5. Ingest a PDF

curl -F "file=@BetterDB_YouTube_Proposal.pdf" \
     -H "X-User-ID: demo" \
     http://localhost:8000/ingest

6. Query

curl -X POST http://localhost:8000/query \
     -H "Content-Type: application/json" \
     -H "X-User-ID: demo" \
     -d '{"query": "What is BetterDB?", "session_id": "default"}'

---

5 BetterDB Features Demonstrated

Feature 1 — MCP Server: Debug in Plain English

Ask Claude Code questions — BetterDB answers from real Redis data:

"What are the slowest commands in the last 24h?"
"Show me memory breakdown by namespace"
"Who are the top clients by command count?"
"Show me any anomalies detected"

Feature 2 — Semantic Cache TTL Bug

After /ingest + /query:

• rag:doc:* — 21 keys, ~31 KB each, w/TTL = 0
• semantic:cache:* — 1+ keys, ~31 KB each, w/TTL = 0

Memory grows unbounded. No CloudWatch alert fires.

Fix: r.expire(cache_key, 604800) — 7 days TTL.

Feature 3 — Agent Memory Runaway

Run multiple queries with the same session_id:

• Key count stays 1 (one HASH key)
• Memory grows per query: 850B → 2KB → 4KB → ...
• w/TTL = 0 — grows forever

Feature 4 — Rate Limiter Burst Detection

Fire 20 parallel requests — queries 11–20 return HTTP 429:

curl ... &
curl ... &
# × 20
wait && echo "done"

See full burst command in step-by-step.md Section 6.

Feature 5 — HGETALL Latency Attribution

Per-key HGETALL (Upstash):     avg=264ms   p95=272ms
Full scan 21 keys (Upstash):   ~6000ms     ← Redis is the bottleneck
Full scan 21 keys (local):     ~5ms        ← 1200× faster

Fix: Pipeline all HGETALLs into one round-trip: ~6000ms → ~300ms.

---

Full Walkthrough

See step-by-step.md for:

• All curl commands
• All Claude Code MCP questions
• Both Upstash and local Valkey setup
• Troubleshooting guide
• Before/after fix comparisons

---

Stack

Component	Choice
API	FastAPI 0.120
LLM	OpenAI gpt-4o-mini
Embeddings	text-embedding-3-small
Redis (cloud)	Upstash Redis
Redis (local)	Valkey 8.1 (Docker)
Observability	BetterDB cloud + agent
MCP	BetterDB MCP → Claude Code
Package manager	uv

---

License

MIT — free to use, modify, distribute.

---

Built by Krish Naik in collaboration with BetterDB

🌐 betterdb.com · 📺 Krish Naik Academy