Certificate

# ===== Skills checklist =====
# After the LangChain track you should be able to:
# [x] Decide LCEL vs agent vs raw LLM call for a feature
# [x] Build a RAG pipeline with chunking, embeddings, retriever, prompt, parser
# [x] Pick the right vector store for your dataset size + cardinality
# [x] Add structured output via with_structured_output(PydanticModel)
# [x] Implement session memory the right way (per request, per user)
# [x] Layer safety controls (input filter, output schema, tool scoping)
# [x] Cache + route + batch to control cost
# [x] Build cheap evals + a CI regression gate
# [x] Wire callbacks for token + cost tracking
# [x] Ship as an API behind rate limits + cost ceiling

# ===== Bookmark =====
# - https://python.langchain.com/docs/get_started
# - https://js.langchain.com
# - https://docs.smith.langchain.com           LangSmith for tracing + evals
# - https://github.com/langchain-ai/langgraph  langgraph for stateful agents
# - https://www.promptingguide.ai              prompt engineering reference

# ===== Portfolio project (8-16 hours) =====
# Build a small RAG product end-to-end:
# 1) Ingest pipeline over a real corpus (docs, support tickets, public dataset)
# 2) Chunk + embed + persist (Chroma / Qdrant / Pinecone)
# 3) RAG chain with structured output
# 4) Eval dataset (50-200 questions) + cheap evaluators in CI
# 5) Rate limiter + cost ceiling
# 6) Caching of identical requests
# 7) FastAPI endpoint + simple web UI
# 8) Logs that capture inputs + outputs (with PII redaction)
# 9) README: architecture diagram, retrieval recipe, eval results, cost analysis

# Bonus:
# - Rerank with a cross-encoder
# - Hybrid search (BM25 + vector)
# - Tool-using agent on top via langgraph
# - LangSmith traces shared publicly

# ===== What 'good' looks like =====
# - Eval pass rate reported HONESTLY (no cherry-picking)
# - Cost per request known + capped
# - Latency p95 within UX budget (< 3s typical)
# - Documented limitations + failure modes
# - Cache hit rate tracked + reported
# - Reproducible: one command rebuilds the index

# ===== Common mistakes to avoid =====
# - No evals (prompt edits regress silently)
# - Free-form output parsed with regex (use structured output)
# - Memory leaked across users / tenants
# - Trusting LLM output as code / SQL / URL without validation
# - Public-internet API without rate limit + cost ceiling
# - Tokens spent on retries instead of caching

# ===== Next steps =====
# - langgraph for stateful, multi-agent workflows
# - Llama.cpp / vLLM for self-hosted inference
# - Embeddings benchmarks: MTEB for picking the right model
# - Voyage AI / Cohere rerank for retrieval quality
# - Anthropic prompt caching / OpenAI batch API for cost wins

# ===== Self-test =====
# If you can:
# 1) Reproduce someone's RAG bug from their prompt + logs
# 2) Reduce a chain's per-call cost by 50% without quality loss
# 3) Ship a production AI feature with a cost ceiling + rate limit + evals
# you have completed the track. Ship the portfolio piece and call it done.

# ===== Track wrap-up =====
# LangChain is the glue, not the magic. The skill that matters is product
# design with LLMs:
# - Use structured output everywhere it makes sense
# - Eval before shipping; gate regressions
# - Cache + route + batch to control cost
# - Treat LLM output as untrusted input to the rest of your system
# Master those and you can build AI features that survive contact with real
# users.

# /certificate/langchain