A policy assistant can have the right documents and still fail to answer a coverage question if retrieval misses the relevant passage. Embeddings help retrieval find content by meaning rather than exact wording, but a poor embedding model can surface irrelevant context or no context at all. section 4.0 section 4.2
An embedding is a vector of numbers representing text as a point in a high-dimensional space, often with 384 or 768 dimensions. A useful model places semantically similar text near each other. Cosine similarity compares vector directions; dot product also takes magnitude into account. section 4.1
Matching different wording
A question about paid adoption leave can be close in embedding space to a policy granting paid parental leave for adoption or foster care. The match depends on the model capturing the relationship despite the different wording. section 4.2
In a retrieval-augmented generation pipeline:
- Load documents, split them into chunks, embed each chunk, and store the resulting vectors for search. section 3.4
- Embed the user’s query in the same vector space and compare it with document vectors to retrieve close matches as context. section 4.0
- Preserve metadata with chunks so results can be filtered by fields such as effective date or policy status. section 4.5.4
from sentence_transformers import SentenceTransformer
import numpy as np
model = SentenceTransformer('pritamdeka/S-PubMedBert-MS-MARCO') #1
chunk_texts = [chunk['content'] for chunk in all_chunks] #2
embeddings = model.encode(chunk_texts, show_progress_bar=True) #3
print(f"Generated embeddings with shape: {embeddings.shape}")
This Python listing encodes health-care policy chunks with a model pretrained on medical literature. section 4.5.4
Choose an embedding model for the domain and retrieval constraints. Generic models can miss specialized vocabulary; in-domain question-document pairs or synthetic questions can support fine-tuning. As new document types are added, embeddings may need retraining. For a vector database, the index dimension must match the model’s output dimension, and the similarity metric must be configured. section 3.7.1 section 4.5.4


















































