Vector databases and Amazon Bedrock Knowledge Bases

Kwame Asante had built databases for years, and he arrived at the sixth AI hour with a doubtful question. "Why do we need a special kind of database?" he asked. "We already have PostgreSQL. I can search for the word 'cancellation' in a few milliseconds."

Noor typed a staff question from the pilot list into the sandbox: "What do I tell someone who wants to move their dog's operation to next week?" The policy that answers it never uses the words "move", "operation" or "next week". It talks about rescheduling surgical procedures with notice. A keyword search for the question's words would find nothing useful. A search by meaning found the right section first.

"That is what vectors are for," Noor said. "And you are half right: we may well keep them in PostgreSQL." This lesson explains how vector search works, which AWS services can store embeddings, and how Amazon Bedrock Knowledge Bases manages the whole RAG pipeline from lesson 5.

Searching by meaning

An embeddings model turns text into a vector, a list of numbers, often hundreds or about a thousand long. Texts with similar meanings get vectors that point in similar directions, even when they share no words. "Move my dog's operation" and "reschedule a surgical procedure" end up close together; "move my dog's food bowl" ends up somewhere else.

To find relevant passages, the system compares the question's vector with the stored vectors and returns the closest ones. A common measure is cosine similarity, which compares the directions of two vectors: 1 means they point the same way, 0 means unrelated, and negative values mean opposite. Other measures, such as Euclidean distance and dot product, are also used; a vector store is configured with the measure that matches the embeddings model.

The code below uses tiny four-number vectors to show the idea. Real embeddings come from a model such as Amazon Titan Text Embeddings, but the arithmetic is the same.

import math

def cosine(a, b):
    dot = sum(x * y for x, y in zip(a, b))
    return dot / (math.sqrt(sum(x * x for x in a)) * math.sqrt(sum(y * y for y in b)))

# Toy 4-dimensional embeddings (real ones have hundreds of dimensions).
chunks = {
    "Rescheduling surgical procedures needs 48 hours notice": [0.9, 0.8, 0.1, 0.0],
    "Late cancellation fees and exceptions":                 [0.7, 0.9, 0.2, 0.1],
    "Payment plans for dental work":                          [0.1, 0.2, 0.9, 0.3],
    "Kennel cleaning between patients":                       [0.0, 0.1, 0.2, 0.9],
}
question = [0.85, 0.7, 0.15, 0.05]   # "move my dog's operation to next week"

ranked = sorted(chunks.items(), key=lambda kv: cosine(question, kv[1]), reverse=True)
for text, vec in ranked:
    print(f"{cosine(question, vec):.3f}  {text}")

0.997  Rescheduling surgical procedures needs 48 hours notice
0.975  Late cancellation fees and exceptions
0.346  Payment plans for dental work
0.141  Kennel cleaning between patients

The two scheduling chunks score far above the others, so they would be retrieved and placed in the prompt. Searching millions of vectors this way, one by one, would be slow, so vector stores use approximate nearest neighbor indexes that find very close matches quickly without comparing every vector.

Keyword, semantic and hybrid search

Vector search is not always better than keyword search. Keywords excel at exact terms: a drug name, a form number such as SOP-114, or a client account number. Semantic search excels at meaning and paraphrase. Many systems combine them in hybrid search, which scores results on both and merges the lists. Some knowledge bases can also rerank results with a second model that reads the question and each candidate passage together and reorders them by relevance, which often improves the top few results at a small extra cost.

Front Desk Helper used hybrid search, because staff often typed SOP numbers and product names, and they also asked questions in everyday words.

AWS services for storing embeddings

A vector database, or vector store, stores embeddings with their text and metadata and answers similarity queries. AWS offers several services that can do this. The exam guide names four examples.

Kwame's instinct was right: if an application already uses PostgreSQL, the pgvector extension on Aurora or RDS can be the simplest choice, because there is no new database to learn or operate. Other services, such as Amazon DocumentDB and Amazon MemoryDB, also offer vector search. Bedrock Knowledge Bases can use several of these as its vector store, as well as some third-party stores.

Amazon Bedrock Knowledge Bases

Building RAG by hand means writing code for loading, chunking, embedding, storing, retrieving and prompting. Amazon Bedrock Knowledge Bases is a managed feature that does that pipeline for you. You point it at your data, choose an embeddings model and a vector store, and it handles ingestion and retrieval. The exam guide gives it as the example of RAG on AWS.

The main parts:

• Data sources. Amazon S3 is the most common; connectors also exist for sources such as web crawlers and document management and collaboration systems. Cascade Ridge kept its SOPs in an S3 bucket owned by the operations team.
• Sync. When documents change, a sync job ingests the changes: parsing the files, chunking them with the strategy you chose, creating embeddings and updating the vector store.
• Embeddings model and vector store. You choose both when creating the knowledge base. The vector index must use the same number of dimensions as the embeddings model, so changing the embeddings model later means re-embedding everything.
• Retrieval APIs. The Retrieve operation returns the most relevant passages with their sources, so your application can build its own prompt. The RetrieveAndGenerate operation retrieves passages and calls a foundation model to produce an answer with citations in one step.
• Metadata and filters. Each document can carry metadata, such as clinic region, audience or effective date, and queries can filter on it, so managers-only documents are retrieved only for managers.
• Guardrails. A guardrail from lesson 3 can be applied to knowledge base answers, including the contextual grounding check that flags answers not supported by the retrieved passages.

# Illustrative request shape for RetrieveAndGenerate (no credentials shown).
# The application sends the question and names the knowledge base and model.
aws bedrock-agent-runtime retrieve-and-generate \
  --input '{"text": "Do we charge a fee if a spay surgery is cancelled the day before?"}' \
  --retrieve-and-generate-configuration '{
      "type": "KNOWLEDGE_BASE",
      "knowledgeBaseConfiguration": {
          "knowledgeBaseId": "KBEXAMPLE01",
          "modelArn": "arn:aws:bedrock:REGION::foundation-model/EXAMPLE-MODEL-ID"
      }
  }'
# The response contains the generated answer and citations pointing to the source chunks.

In June 2026 AWS also launched Amazon Bedrock Managed Knowledge Base, a fully managed option that looks after the vector storage itself as well as ingestion, with built-in hybrid search, ranking and agentic retrieval that plans multi-step searches for complex questions. It removes the choice of vector store entirely, at the cost of less control over it. At launch it was offered in a short list of Regions that included US West (Oregon), so the Denver team noted it as a simpler alternative to compare with a standard knowledge base before the pilot. Either way, the RAG ideas in this lesson apply unchanged.

Front Desk Helper used Retrieve rather than RetrieveAndGenerate, because the team wanted its own tested prompt template from Prompt Management, with the refusal rules and short format from lessons 2 and 3. Night Notes did not use a knowledge base at all: everything it needed was in the call note.

Costs and operations

A knowledge base has several cost components, and Lorraine Beaudry wanted each one listed: embedding the documents during ingestion and again whenever they change; the vector store itself, which runs and is billed whether or not anyone asks a question; embedding each question; any reranking; and the foundation model tokens for each answer. For 140 documents, embedding costs were tiny. The vector store was the largest fixed cost, which is why the team compared a serverless option with simply adding pgvector to an Aurora database that already existed for the booking system.

Operations matter too. Someone must own the documents, run or schedule syncs after changes, and check that retired documents are removed. Hannah added a quarterly review: the operations team confirms that every SOP in the bucket is current, and the knowledge base is re-synced afterwards.

Choosing an embeddings model

The embeddings model matters as much as the vector store. It decides which texts count as similar, so a model that handles the vocabulary of the documents well gives better retrieval. Bedrock offers embeddings models from Amazon and other providers, and the Amazon Titan Text Embeddings models let you choose the vector size; smaller vectors cost less to store and search, larger ones can capture more detail. Multilingual embeddings models place a question and a passage close together even when they are written in different languages, which would matter if clinic staff searched Spanish documents with English questions. The vector store's index must be created with the same number of dimensions as the chosen model.

Jonah tested two embeddings models on the 200-question test set by measuring how often the right SOP section appeared in the top four results. The difference was small, so the team chose the cheaper option and recorded the result so the decision could be revisited if the document collection grew.

Retrieval settings are tuned with evaluation, not instinct. The main ones are how many passages to return, whether to use hybrid search, whether to rerank, and any minimum similarity score below which a passage is ignored. Each change can help one kind of question and hurt another, so the team changed one setting at a time and reran the 200-question test set after each. The results, kept in a simple table, became the record of why the knowledge base was configured the way it was.

Kwame left the session with a new item on his list: "Try pgvector on the BookVet Aurora cluster in the sandbox." He also admitted that the operation and rescheduling demonstration had been convincing.