Normalization and Scaling — Guardrails and Fix Pattern

Use this page when vector similarity is unstable because embeddings are not normalized or scaling factors differ between training and retrieval.
This failure often appears when cosine distance is requested but vectors are stored raw, or when IP/dot metrics exaggerate magnitude.

Open these first

Visual map and recovery: RAG Architecture & Recovery
Embedding vs meaning: embedding-vs-semantic.md
Metric mismatch: metric_mismatch.md
Chunking checklist: chunking-checklist.md

Core acceptance

Vectors are L2-normalized when using cosine similarity.
ΔS(question, retrieved) ≤ 0.45, stable across three paraphrases.
Coverage ≥ 0.70 on the target section.
λ remains convergent across seeds.

Typical breakpoints and the right fix

Cosine similarity reported but vectors not normalized
→ metric_mismatch.md
Dot product used without rescaling (large norm vectors dominate retrieval)
→ Normalize or rescale embeddings before indexing.
Cross-model mixing (embeddings from different checkpoints with different norms)
→ Re-normalize the corpus and queries to unit length.
Hybrid dense + sparse weighting unstable (scale mismatch between BM25 scores and vector norms)
→ Apply explicit min-max or z-score scaling before weighted sum.

Fix in 60 seconds

Check norms
Sample 100 embeddings. Compute mean L2 norm. If not ~1.0 under cosine, normalization missing.
Normalize queries
Ensure query_vector = vector / ||vector|| before retrieval when using cosine.
Corpus re-index
Drop and rebuild index with normalized vectors if store does not enforce it.
Hybrid scaling
Normalize dense similarity scores into the same 0–1 range as BM25 before combining.

Copy-paste probe

import numpy as np

def check_norms(vectors):
    norms = np.linalg.norm(vectors, axis=1)
    return norms.mean(), norms.std()

mean_norm, std_norm = check_norms(sample_vectors)
print("Mean norm:", mean_norm, "Std:", std_norm)

Target: mean ≈ 1.0, std ≤ 0.05 for cosine retrieval.

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

🧭 Explore More

Module	Description	Link
WFGY Core	WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack	View →
Problem Map 1.0	Initial 16-mode diagnostic and symbolic fix framework	View →
Problem Map 2.0	RAG-focused failure tree, modular fixes, and pipelines	View →
Semantic Clinic Index	Expanded failure catalog: prompt injection, memory bugs, logic drift	View →
Semantic Blueprint	Layer-based symbolic reasoning & semantic modulations	View →
Benchmark vs GPT-5	Stress test GPT-5 with full WFGY reasoning suite	View →
🧙‍♂️ Starter Village 🏡	New here? Lost in symbols? Click here and let the wizard guide you through	Start →

👑 Early Stargazers: See the Hall of Fame — Engineers, hackers, and open source builders who supported WFGY from day one.

⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.

7.2 KiB Raw Blame History Unescape Escape