Can I vibecode Sourcegraph Cody?

Build the closest honest consolation tool inspired by Sourcegraph Cody, focusing on repository-wide context understanding.
Use exactly this stack: Python 3.12, Typer for the CLI, SQLite for local embeddings/metadata, and a local vector search library (like FAISS or simply SQLite vector extensions); do not offer alternative stacks.
Primary job: Build a repository-local AI code assistant that indexes an entire local codebase, generates embeddings for functions and classes, retrieves relevant context based on user queries, calls a user-provided LLM, proposes diffs, and records interactions.
Start from an empty folder and create the complete working project. Make the default mode single-user and private.
Store all embeddings, file metadata, and chat history locally in SQLite. Put every secret and external API credential in .env and provide .env.example.
Implement a command to index the current directory, ignoring files based on .gitignore, and splitting code into semantic chunks.
Implement a chat interface in the CLI where the user can ask questions about the codebase, and the tool fetches the top-K relevant code snippets to construct a highly contextualized prompt.
Include clear empty, loading (during indexing and API calls), validation, success, and failure states.
Provide commands to clear the index, re-index, and export chat history.
Validate untrusted input and never log secrets or transmit private file contents to unauthorized endpoints.
Deliberately exclude these paid-product advantages: enterprise SSO, policy enforcement, managed cloud embeddings, and massive-scale multi-repository indexing.
Write focused unit tests for the chunking, indexing, and retrieval logic. Add one end-to-end smoke test that proves the core index-and-query loop works on a dummy python project.
Create a README with setup, architecture, embeddings configuration, data location, and limitations.
Finish by running the tests and listing the exact commands used.