Study

A retrieval augmented generation (RAG) app.

User Interface

Flows

There are two flows within the program. One that happens at program start time, and another that happens for each user query submitted through the UI.

Diagram

At Startup

When the program starts up, each local file has an embedding generated for its entire contents. This prepares local content for similarity scoring against user queries.

• Program starts
• Local fact directory is scanned for files (assumed .txt)
• Each file has its embedding generated and stored in-memory
• UI starts, user can submit a query

On User Query

When a user submits a query via the UI, the user-supplied query is expaned. The original query and all expanded queries are compared to facts and a cited response is generated. The facts cited in the response are checked against locally stored facts. Finally, the response is rendered in the UI.

• User submits a query (ex: "why is the sky blue?")
• The query is expanded into multiple related queries
• An embedding is generated for each query (original + expanded queries)
• Each query embedding is matched with similar in-memory facts
• Queries and facts are used to prompt for a cited response
• Facts referenced in the cited response are checked (existence only)
• Fact-checked response is provided to the user (free-form commentary + references)

Configuration

Configuration is handled through environment variables:

• OPENAI_API_KEY (required)
• OPENAI_LANGUAGE_MODEL (default: gpt-4o)
• OPENAI_EMBEDDING_MODEL (default: text-embedding-3-large)

Limitations

• Fact storage and search is inefficient
• Fact storage does not persist
• Fact storage is not continuous
• Fact checking does not check meaning, only existence of files referenced
• User experience is a single query/response, no persistence or threading

Implementation

• Single-file, sub-1k LoC Go program main.go
• Configuration via environment variables run.bat
• No dependencies except the Go standard library