Skill Improvement

Use Python variables, basic data types, and type conversions to represent application data.

Control a Python program with conditional statements.

Iterate over lists and dictionaries with Python for-loops.

Define and call Python functions with parameters and return values.

Extract and update values in nested Python dictionaries and lists.

Read and write UTF-8 text files containing English and Japanese characters.

Load and save structured data in JSON format.

Use a Python traceback to correct common syntax, type, and key errors.

Navigate folders and run Python programs from a command-line terminal.

Create a Python virtual environment and install packages within it.

Record and push a small Python project using Git and GitHub.

Load an API key from an environment variable without exposing it in source code.

Identify the method, URL, headers, body, and status code in an HTTP exchange.

Send GET and POST requests to a REST API from Python.

Extract required fields from a nested JSON API response.

Handle API timeouts, unsuccessful status codes, and rate-limit responses.

Distinguish an LLM, an inference service, an API, and a client application.

Explain how next-token prediction produces a text response during inference.

Relate tokens and context-window limits to the size of an LLM request.

Call a free-tier hosted LLM from Python and extract its generated response.

Use system instructions and user messages to define an LLM task clearly.

Compare how decoding settings affect the consistency and variety of responses.

Prompt an LLM to return data that follows a specified JSON structure.

Test a prompt against representative examples and document its failure cases.

Define the users, supported questions, and success criteria for a Japan internship assistant.

Design a prompt that extracts structured internship details from supplied text.

Create examples that test the prompt on both English and Japanese internship text.

Validate an LLM's structured internship output before using it in an application.

Build a Streamlit interface that accepts a question and displays a response.

Connect a Streamlit interface to an LLM API call.

Record useful diagnostic information without logging secrets or sensitive user data.

Evaluate the prototype against its stated success criteria.

Explain how retrieval-augmented generation differs from prompting an LLM with no external source.

Select authoritative internship documents that may be used lawfully in the project.

Clean extracted English and Japanese document text while preserving its meaning.

Split bilingual documents into retrieval chunks using explicit size and overlap rules.

Attach source, language, date, and URL metadata to each document chunk.

Explain how an embedding represents semantic similarity between pieces of text.

Generate multilingual embeddings for document chunks with a free-to-use model.

Store document embeddings and metadata in a searchable vector index.

Retrieve the top-k document chunks for an English or Japanese query.

Restrict retrieval results using document metadata filters.

Calculate cosine similarity for a small pair of example vectors.

Create a bilingual set of test questions with known relevant source chunks.

Calculate recall at k for retrieval results on the test set.

Compare retrieval quality across two chunk-size configurations.

Diagnose a retrieval failure by inspecting queries, chunks, scores, and metadata.

Package the selected retrieval configuration as a reusable Python function.

Construct an LLM prompt that clearly separates retrieved evidence from user instructions.

Generate an answer that cites the source metadata of its supporting chunks.

Make the assistant abstain when retrieved evidence is insufficient.

Defend the RAG prompt against instructions embedded inside retrieved documents.

Connect document retrieval and grounded answer generation in the Streamlit application.

Create a rubric for grading answer relevance, evidence support, and citation correctness.

Evaluate the RAG application on the bilingual test-question set.

Use evaluation results to select and justify one RAG configuration.

Distinguish a fixed LLM workflow, a tool-using workflow, and an autonomous agent.

Describe a structured tool interface for searching the project's internship records.

Implement one bounded tool-using workflow with validated inputs and an execution limit.

Test the tool-using workflow for invalid requests and unsupported actions.

Separate application configuration, retrieval, generation, and interface code into modules.

Add caching and request limits to control latency and API usage.

Deploy the application using a free-tier hosting option while keeping secrets private.

Present the project with a GitHub README covering setup, architecture, evaluation results, limitations, and a demo.