Context Injection & Vector RAG

Generative models do not know what your local workspace looks like unless you feed it to them. Retrieval-Augmented Generation (RAG) is the technique of searching your repository and injecting relevant file contents into the prompt.

---

Managing the Context Window

Modern workspace tools automatically compile vector index profiles of your project. However, index size can quickly balloon, causing model memory issues.

[ Large Codebase ] ──> [ Vector Embeddings Index ] ──> [ Context Selection ]
                                                                 │
                                                                 ▼
    [ AI Chat Prompts ] <──────────────────────────────── [ Match Injector ]

1. Vector Embeddings

2. Context Pruning

• Exclude large media directories and database dumps.
• Hide compilation targets (build/, .next/, dist/) using project ignores.

Playbook: Prompt Tagging

Instead of feeding the entire codebase, use target reference keywords:

• Tag the database schema (schema.prisma).
• Tag the API interface definition file.
• Tag the components layout definitions.

Checklist: Context Optimization

• [ ] Exclude npm build targets from vector index scanning configurations.
• [ ] Tag files manually using context tools (@ references) rather than pasting full directories.
• [ ] Ensure local database seed dumps are kept out of active chat buffers.
• [ ] Rebuild vector indices after making major database schema changes.
• [ ] Check token usage indicators to avoid context window overflows.