Homework

Task 1: Add Conversation Memory to LangChain

Goal: Extend your LangChain chain from the lab to support multi-turn conversations with memory.

Instructions

Build a conversational chain that remembers previous exchanges:

1. Create a ChatPromptTemplate that includes a placeholder for conversation history:

   from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
   
   prompt = ChatPromptTemplate.from_messages([
       ("system", "You are a helpful technical assistant. Be concise."),
       MessagesPlaceholder(variable_name="history"),
       ("human", "{input}")
   ])

2. Implement a conversation loop that maintains a list of messages and passes them as the history variable on each turn.
3. Have at least a 5-turn conversation where later questions depend on context from earlier turns. For example:
   • Turn 1: "What is a vector database?"
   • Turn 2: "Give me three examples of popular ones."
   • Turn 3: "Which of those would you recommend for a small project?"
   • Turn 4: "How would I set it up with Python?"
   • Turn 5: "Can you summarise what we discussed?"
4. Verify that the model correctly references information from earlier turns in its later responses.

Self-check

• ✓ Your conversation loop runs for at least 5 turns.
• ✓ The model's answer to turn 3 correctly references the examples from turn 2.
• ✓ The turn 5 summary covers topics from the entire conversation.
• ✓ You can explain how the history is passed to the model on each turn.

Task 2: Swap LLM Providers

Goal: Experience the provider-agnostic benefit of frameworks by switching from Claude to another provider.

Instructions

Take your working LangChain chain (from the lab or Task 1) and your LlamaIndex RAG pipeline:

1. Swap the LLM from Claude to OpenAI (or vice versa, or to a local model via Ollama if you have one running):

   # LangChain: change one import and one line
   from langchain_openai import ChatOpenAI
   model = ChatOpenAI(model="gpt-4o", temperature=0)
   
   # LlamaIndex: change one import and one line
   from llama_index.llms.openai import OpenAI
   Settings.llm = OpenAI(model="gpt-4o")

2. Run the exact same queries on both providers.
3. Compare the outputs:
   • Are the answers substantively different?
   • Is one more verbose or more concise?
   • For the RAG pipeline: does one cite sources more accurately?
4. Record your observations in a short comparison (~150 words).

Self-check

• ✓ You ran the same queries on at least two different LLM providers.
• ✓ The provider swap required changing no more than 2–3 lines of code.
• ✓ Your comparison notes specific differences, not just "they were similar".
• ✓ You can articulate one scenario where you would prefer each provider.

Task 3: Framework Decision Memo

Goal: Practice the real-world skill of choosing the right tool for a project.

Instructions

Choose one of the following project scenarios (or invent your own):

• Scenario A: A customer support chatbot that answers questions about your product documentation (50 pages of docs, ~100 queries/day).
• Scenario B: A code review assistant that takes a git diff and produces a structured review with suggestions (single LLM call per review, ~20 reviews/day).
• Scenario C: An internal research tool that lets employees query 10,000 internal documents, with follow-up questions and the ability to search the web when internal docs are insufficient.

Write a decision memo (300–500 words) that recommends one of these approaches:

1. Direct API calls only
2. LangChain
3. LlamaIndex
4. A combination

Your memo should address:

• Why the chosen approach is the best fit for the scenario's requirements
• What trade-offs you are accepting
• What would need to change about the project for you to choose a different approach
• Estimated dependency footprint and maintenance burden

Self-check

• ✓ Your memo is 300–500 words.
• ✓ You make a clear recommendation with specific reasoning (not "it depends").
• ✓ You address at least two trade-offs of your chosen approach.
• ✓ You describe a concrete condition that would change your recommendation.
• ✓ A technical manager could use your memo to make a decision.