| # MINIMIZATION PROOF | |
| ## Source | |
| | | Lines | | |
| |-|-------| | |
| | Leader source (`dspy/retrievers/embeddings.py`) | 261 | | |
| | `kernel.py` | 9 | | |
| ## LOC Ratio | |
| ``` | |
| 9 / 261 = 0.0345 → 3.45% of source absorbed | |
| ``` | |
| Reduction factor: **29×** | |
| ## What was dropped | |
| - FAISS index construction and approximate search (~80 lines) | |
| - Unbatchify / async batching wrapper (~40 lines) | |
| - DSPy `Prediction` return type and dspy import chain (~30 lines) | |
| - `__init__` configuration surface area (~25 lines) | |
| - Docstrings, type annotations, error handling (~50 lines) | |
| - All external dependency guards (~20 lines) | |
| ## What was kept (core pattern) | |
| 1. `np.random.default_rng(seed)` — deterministic query vector (stand-in for real encoder) | |
| 2. Cosine similarity via `dot / (norm * norm)` — identical to DSPy normalize-then-dot pattern | |
| 3. `sorted(..., reverse=True)[:k]` — top-k selection | |
| 4. Dual-store pass: PIRWA features first, QILLQA codex priors (capped at 8) second | |