LangChain-Step-By-Step

This repository is a step-by-step onboarding assignment for learning how to build LLM applications with LangChain and OpenAI. Interns should implement step_1.py through step_6.py in order and finish with a small AI assistant that can handle both external RAG dataset lookup and general conversation.

This assignment assumes Python 3.11+ or higher, LangChain v1, and an OpenAI API key. For Step 2, the assignment uses the more modern ChatPromptTemplate + MessagesPlaceholder pattern instead of the older ConversationChain style.

Learning Goals

• Safely manage .env files and API keys
• Run basic LLM calls with ChatOpenAI
• Feed conversation history back into prompts
• Build a simple RAG workflow using a Hugging Face dataset
• Structure LLM outputs with Pydantic
• Let an agent decide when to use a tool

Repository Structure

.
├── data/
│   └── questions.txt
├── docs/
│   └── reviewer-rubric.md
├── outputs/
│   └── .gitkeep
├── .env.example
├── requirements.txt
├── step_1.py
├── step_2.py
├── step_3.py
├── step_4.py
├── step_5.py
└── step_6.py

Getting Started

1. Fork or clone this repository.
2. Create a virtual environment (you can set up your development environment in any way you prefer). Make sure to use Python 3.11 or higher.

   python3.11 -m venv .venv
   source .venv/bin/activate

   To deactivate the virtual environment later, run:

   deactivate

3. Install dependencies.

   pip install -r requirements.txt

4. Copy .env.example to .env and fill in OPENAI_API_KEY.

   cp .env.example .env

5. Implement the assignment in order, starting from step_1.py.

Assignment Rules

• Each step should extend the result of the previous step.
• If you get stuck, read the official docs first.
• Open a Pull Request after completing Step 6.

Step Overview

Step	Topic	Goal	What to Check
1	Basic Setup	Read questions.txt, generate answers, and save them to outputs/step_1_output.txt	API integration, file I/O
2	Conversation Memory	Build a chatbot that remembers earlier turns	Name/context retention
3	Structured Output	Return intent and answer as JSON/Pydantic	Output parsing, validation
4	Simple RAG	Build dataset-based QA with neural-bridge/rag-dataset-12000	Dataset loading, Vector DB insertion, context retrieval, answer generation
5	Final Agent	Build an assistant that connects with the Vector DB to answer dataset questions	Tools, agent flow, reuse
6	Language Filter	Detect language and refuse to answer if not Japanese or English	Guardrails, prompt engineering or routing

---

Step 1 — First Conversation with an LLM

Mission

• Read data/questions.txt line by line.
• Send each question to ChatOpenAI.
• Save the questions and answers in a readable format to outputs/step_1_output.txt.

Required

• Read the API key from .env
• Ignore empty lines
• Overwrite the output file on each run

---

Step 2 — A Chatbot with Memory

Mission

• Build a chatbot that can continue a multi-turn conversation.
• The following flow should work:
  • My name is Jimin.
  • What is my name?

Required

• Use MessagesPlaceholder
• Choose either full-history memory or a recent-window memory approach
• Manually verify at least 3 turns of conversation

Hint

• In modern LangChain flows, ChatPromptTemplate + MessagesPlaceholder is more instructive than the legacy ConversationChain pattern.
• Managing history as a list is perfectly fine for this assignment.

---

Step 3 — Structured Output

Mission

• Analyze user input and return it in the following format.

{
  "intent": "question | greeting | request",
  "answer": "..."
}

Required

• Use PydanticOutputParser or an equivalent explicit schema validation flow
• Handle parse failures with retries or explicit error handling

---

Step 4 — External Data with Simple RAG

Mission

• Use the Hugging Face dataset neural-bridge/rag-dataset-12000 to build a simple RAG pipeline.
• Use the dataset's context, question, and answer fields to implement a retrieval + answer generation flow.
• At minimum, run retrieval-based answering on a few test questions and compare your output with the provided answers.

Required

• Load neural-bridge/rag-dataset-12000 with the datasets library
• Use the train split context field as the retrieval corpus
• Pick a few test split questions and run retrieval + generation on them
• Only place relevant context or chunks into the prompt

Vector DB Implementation (Required)

• A Vector DB (e.g., Chroma) is required for retrieval.
• Extract the context from exactly train[:200] and convert them into Document objects.
• Create embeddings and store them in a Vector DB collection.
• If the retrieved context is not enough to answer, say that the answer could not be verified from the context.

Suggested Implementation Order

1. Load the dataset with load_dataset()
2. Slice the dataset to train[:200] and test[:20]
3. Convert train[:200] contexts into Document objects and store them in a Vector DB via embeddings
4. Retrieve relevant context using similarity search for each test question
5. Insert the retrieved context into the prompt and generate an answer
6. Compare your result with the dataset answer or log the differences

Short Chroma Guide

• Install these packages in your local Python environment:

  pip install langchain-chroma chromadb

• Use persist_directory if you want a local persistent DB.
• If you want to try a local server flow, install the Chroma CLI and run chroma run.
• For installation details and the latest setup instructions, refer to the official Chroma website and docs:
  • https://www.trychroma.com/
  • https://docs.trychroma.com/

---

Step 5 — Autonomous Assistant

Mission

• Integrate with the Vector DB from Step 4.
• If a user question needs external knowledge retrieval, use the Vector DB lookup tool.
• If it is general conversation, let the LLM answer directly.
• In other words, let the LLM decide whether it should use the retrieval tool.

Required

• Define at least one tool
• Make the agent's reasoning or execution log inspectable
• Demo both of the following:
  • one question selected from the dataset
  • one general conversation prompt

---

Step 6 — Language Routing / Filtering

Mission

• Add a guardrail that detects the language of the user's input before answering.
• If the question is not in Japanese (日本語) or English (English), the assistant must decline to answer (e.g., "I can only answer in Japanese or English.").
• This rule applies to both general conversation and dataset questions.

Required

• Use LangChain routing mechanisms or direct prompt engineering to filter based on language.
• Demo an allowed language query and a blocked language query (e.g., Korean, Spanish).

---

Submission Guide

• Open a Pull Request to track and submit your progress.
• In your PR description, briefly include what you implemented in each step.
• AI Coding Tools: You are completely free to use AI coding tools (such as Antigravity, Claude Code, etc.) for this assignment. If you need a license or support for any specific AI coding tool, please let us know and we will provide it for you.