Vectors embedding is used to construct semantic meaning in datapoints, allowing search engines, data models and AI algorithms to find information that is similar to each other. This technique is useful in making sure that queries also find information / datapoints that are close to each other, rather than needing exact input from the user.
Whilst vector embeddings indeed allow for semantic searching, in practice users commonly need a combination of soft and hard datapoints. For example, requesting “find policy documents relating to the first project in 2026” will return every bit of information relating to that first project, whilst the user might only need 1 or 2 specific policies. We often recommend that vector embedding is applied to specific bits of information dependent on the workflow, rather than encoding the entire dataset.
Vector embeddings turn unstructured text into fixed-length numeric coordinates. Distance in that space approximates semantic relatedness, which is why a query can retrieve a passage that never uses the same words. Approximate Nearest Neighbor indexes — commonly HNSW graphs or IVF clusters — avoid brute-force scans so retrieval stays fast as the corpus grows. Quality hinges on chunk boundaries, embedding model choice, normalization, metadata filters, and re-index when content or the model changes. Treat ANN parameters (efSearch, nprobe, candidate count) as measured knobs on recall@k vs latency, not set-and-forget defaults.
# Embed + HNSW-style ANN query (illustrative; swap client for your vector DB)
from openai import OpenAI
import numpy as np
client = OpenAI()
def embed(texts: list[str]) -> np.ndarray:
res = client.embeddings.create(model="text-embedding-3-small", input=texts)
return np.array([d.embedding for d in res.data], dtype=np.float32)
# Assume `index` is an ANN index (e.g. hnswlib / Qdrant / pgvector) over corpus embeddings
query = "How do we reduce citation gaps in AI answers?"
q = embed([query])[0]
# ids, scores = index.knn_query(q, k=10) # tune k + search_ef for recall vs latency