Define the query prefix and doc format before the snippet uses them

README.md

@@ -36,7 +36,11 @@ which reranks this model's top candidates.
 
 ## Usage
 
-Embeddings are L2-normalized, so cosine similarity is a plain dot product.
+The two sides are encoded asymmetrically — a task description carries an
+instruction prefix, a skill is the bare `name | description | body`
+concatenation. Encode them the way the model was trained or retrieval quality
+drops. Embeddings come back L2-normalized, so cosine similarity is a plain dot
+product.
 
 ```python
 import torch
@@ -47,6 +51,15 @@ MODEL = "EverMind-AI/skillcorpus-embedding-0.6b"
 tok = AutoTokenizer.from_pretrained(MODEL, padding_side="left")
 model = AutoModel.from_pretrained(MODEL, dtype=torch.bfloat16).cuda().eval()
 
+QUERY_INSTRUCTION = (
+    "Instruct: Given a task description, retrieve the most relevant "
+    "skill document that would help an agent complete the task\nQuery:"
+)
+
+
+def doc(name, description, body):
+    return f"{name} | {description} | {body}"
+
 
 def last_token_pool(hidden, attention_mask):
     if attention_mask[:, -1].sum() == attention_mask.shape[0]:   # left padding
@@ -70,24 +83,6 @@ print((query @ docs.T).tolist())
 # -> [[0.74, 0.06]]
 ```
 
-### Input format
-
-The two sides are asymmetric — encode them the way the model was trained or
-retrieval quality drops:
-
-```python
-QUERY_INSTRUCTION = (
-    "Instruct: Given a task description, retrieve the most relevant "
-    "skill document that would help an agent complete the task\nQuery:"
-)
-
-def doc(name, description, body):
-    return f"{name} | {description} | {body}"
-```
-
-Queries carry the instruction prefix; skill documents are the bare
-`name | description | body` concatenation, no prefix.
-
 ## Intended use
 
 First-stage retrieval over a large skill registry: encode the registry offline,