Glossary

Common terms and definitions used in the NorthBuilt RAG System documentation.

Table of contents

  1. A
    1. API Gateway
    2. ARN (Amazon Resource Name)
    3. AWS Bedrock
  2. B
    1. Bootstrap
  3. C
    1. Claude Sonnet 4.5
    2. CloudFront
    3. CloudWatch
    4. Cognito
    5. CORS (Cross-Origin Resource Sharing)
    6. Cosine Similarity
  4. D
    1. DynamoDB
  5. E
    1. Embedding
  6. F
    1. Foundation Model
  7. G
    1. GitHub Actions
    2. GitHub Pages
  8. H
    1. HTTP API
  9. I
    1. IAM (Identity and Access Management)
    2. IaC (Infrastructure as Code)
  10. J
    1. Jekyll
    2. JWT (JSON Web Token)
  11. K
    1. Knowledge Base
  12. L
    1. Lambda
    2. Lambda Cold Start
  13. M
    1. MAU (Monthly Active Users)
  14. N
    1. Namespace
  15. O
    1. OAuth 2.0
    2. OIDC (OpenID Connect)
    3. OpenSearch Serverless
  16. P
    1. Pinecone
    2. PITR (Point-in-Time Recovery)
  17. Q
    1. Query
    2. Query Lambda
  18. R
    1. RAG (Retrieval-Augmented Generation)
    2. Reserved Concurrency
    3. Route53
  19. S
    1. S3 (Simple Storage Service)
    2. Secrets Manager
    3. Semantic Search
    4. Serverless
  20. T
    1. Terraform
    2. Throttling
    3. Titan Embeddings V2
  21. U
    1. Upsert
  22. V
    1. Vector
    2. Vector Database
    3. VPC (Virtual Private Cloud)
  23. S (Additional)
    1. S3 Vectors
  24. W
    1. Webhook
  25. X
    1. X-Ray
  26. Acronyms

A

API Gateway

AWS service that creates and manages HTTP APIs. We use HTTP API (not REST API) for lower cost and better performance.

ARN (Amazon Resource Name)

Unique identifier for AWS resources. Format: arn:aws:service:region:account-id:resource-type/resource-id

AWS Bedrock

AWS managed service for AI foundation models (documentation). We use it for:

  • Claude Sonnet 4.5: Response generation
  • Titan Embeddings V2: Vector embeddings

B

Bootstrap

The process of setting up foundational AWS resources before Terraform can deploy:

  • S3 bucket for Terraform state
  • DynamoDB table for state locking
  • GitHub OIDC provider
  • IAM role for GitHub Actions

C

Claude Sonnet 4.5

Anthropic’s AI model for natural language generation. We use it via AWS Bedrock for generating responses from retrieved context.

Model ID: us.anthropic.claude-sonnet-4-5-20250929-v1:0 Cost: $3/M input tokens, $15/M output tokens

CloudFront

AWS Content Delivery Network (CDN) that distributes the web UI globally for low latency.

CloudWatch

AWS monitoring service providing logs, metrics, and alarms for the system.

Cognito

AWS authentication service. We use it for user management and Google OAuth integration.

CORS (Cross-Origin Resource Sharing)

Browser security mechanism allowing the web UI to call the API Gateway from a different domain.

Cosine Similarity

Mathematical measure of similarity between two vectors, used by S3 Vectors to find relevant documents. Range: -1 to 1 (1 = identical, 0 = unrelated, -1 = opposite).

D

DynamoDB

AWS NoSQL database. We use it for:

  • Terraform state locking
  • Document classification results

E

Embedding

Numerical vector representation of text that captures semantic meaning. Generated by Titan Embeddings V2 (1024 dimensions).

Example: "serverless architecture"[0.23, -0.45, 0.67, ...] (1024 numbers)

F

Foundation Model

Pre-trained AI model available on AWS Bedrock. We use:

  • Claude Sonnet 4.5 for text generation
  • Titan Embeddings V2 for embeddings

G

GitHub Actions

CI/CD platform integrated with GitHub. We use it for automated Terraform deployments via OIDC.

GitHub Pages

Static site hosting provided by GitHub. Hosts our documentation at https://craftcodery.github.io/compass

H

HTTP API

Type of API Gateway that’s simpler and cheaper than REST API. Lacks some features (caching, usage plans) but 70% cheaper.

I

IAM (Identity and Access Management)

AWS service for managing permissions. We use it for Lambda execution roles with least-privilege policies.

IaC (Infrastructure as Code)

Managing infrastructure through code (Terraform) rather than manual configuration.

J

Jekyll

Static site generator used by GitHub Pages to build documentation websites from Markdown files.

JWT (JSON Web Token)

Token format used by Cognito for authentication. Contains user identity and is cryptographically signed.

K

Knowledge Base

Collection of documents that can be searched semantically (documentation). In our system, documents are stored as vectors in S3 Vectors via Bedrock Knowledge Base.

L

Lambda

AWS serverless compute service. We have multiple Lambda functions:

  • Chat: Main orchestrator (query understanding, Bedrock KB retrieval, response generation)
  • Classify: Document classification via domain mapping
  • Ingest: Document ingestion to S3/Bedrock KB
  • Webhooks: Fathom, HelpScout, Linear (3 handlers)
  • Sync: Fathom, HelpScout, Linear (3 sync functions with workers)

Lambda Cold Start

Delay (~1 second) when Lambda function is invoked for the first time or after being idle. Can be eliminated with provisioned concurrency.

M

MAU (Monthly Active Users)

Cognito pricing metric. First 50K MAU are free.

N

Namespace

Logical partition within a vector index. Used for multi-tenancy isolation.

O

OAuth 2.0

Industry-standard protocol for authorization. We use Google OAuth for user authentication via Cognito.

OIDC (OpenID Connect)

Authentication layer on top of OAuth 2.0. GitHub Actions uses OIDC to authenticate with AWS without long-lived credentials.

OpenSearch Serverless

AWS vector database we used initially. Replaced with S3 Vectors for cost savings and simpler architecture.

P

Pinecone

Third-party managed vector database. Previously used in our system (see ADR-010). Replaced with S3 Vectors for native AWS integration.

PITR (Point-in-Time Recovery)

DynamoDB feature allowing restore to any point in last 35 days. Enabled on classify table.

Q

Query

User question submitted to the chat API. Example: “What is the system architecture?”

Query Lambda

Deprecated. Document retrieval is now handled directly by the Chat Lambda via Bedrock Knowledge Base APIs.

R

RAG (Retrieval-Augmented Generation)

AI technique combining retrieval (find relevant documents) with generation (create response using documents as context).

Flow: User query → Generate embedding → Search vectors → Retrieve documents → Generate response

Reserved Concurrency

Number of Lambda instances reserved for a function, preventing runaway costs from infinite loops.

Route53

AWS DNS service. Can be used for custom domain setup (optional).

S

S3 (Simple Storage Service)

AWS object storage. We use it for:

  • Terraform state file
  • Web UI static assets

Secrets Manager

AWS service for storing sensitive data (API keys, passwords). We store:

  • Fathom, HelpScout, Linear API keys
  • Google OAuth client secret

Search based on meaning rather than keyword matching. Uses vector embeddings and similarity measures.

Example: Query “serverless compute” matches documents about “Lambda” even though “serverless” isn’t in the text.

Serverless

Architecture pattern where cloud provider manages servers. Benefits: auto-scaling, pay-per-use, no server management.

T

Terraform

Infrastructure as Code tool for provisioning AWS resources. We use it to deploy 100% of our infrastructure.

Throttling

Rate limiting mechanism in API Gateway to prevent abuse. Current limits: 10 req/s, 20 burst.

Titan Embeddings V2

AWS Bedrock model for generating text embeddings (1024 dimensions) (documentation).

Model ID: amazon.titan-embed-text-v2:0 Cost: $0.0001 per 1000 tokens

U

Upsert

Vector database operation to insert or update vectors. “Upsert” = insert + update.

V

Vector

Array of numbers representing text semantically. Example: 1024-dimensional vector from Titan.

Vector Database

Database optimized for storing and querying high-dimensional vectors. We use S3 Vectors via Bedrock Knowledge Base.

VPC (Virtual Private Cloud)

AWS isolated network. Not used in our architecture for simplicity.

S (Additional)

S3 Vectors

AWS purpose-built vector storage for Bedrock Knowledge Base (documentation). Native integration with zero infrastructure management.

Configuration:

  • Dimension: 1024 (matches Titan embeddings)
  • Metric: Cosine similarity
  • Storage: Fully managed by AWS
  • Non-filterable keys: AMAZON_BEDROCK_TEXT, AMAZON_BEDROCK_METADATA (required for 100% ingestion)
  • Data deletion policy: DELETE (vectors auto-removed when S3 docs deleted)

Important: LLM parsing is disabled. Metadata is provided via sidecar .metadata.json files to avoid the 2KB filterable metadata limit (metadata filtering).

W

Webhook

HTTP callback mechanism where external service sends data to our API when events occur.

We support webhooks from:

  • Fathom: Video transcripts
  • HelpScout: Support tickets
  • Linear: Teams (clients) for entity registry; projects synced for display purposes

X

X-Ray

AWS distributed tracing service. Can be enabled on Lambda functions for performance debugging (optional).


Acronyms

Acronym Full Form Description
ACM AWS Certificate Manager SSL/TLS certificates
ADR Architecture Decision Record Documentation of design decisions
API Application Programming Interface HTTP endpoints
ARN Amazon Resource Name AWS resource identifier
CDN Content Delivery Network CloudFront
CORS Cross-Origin Resource Sharing Browser security
HTTP Hypertext Transfer Protocol Web protocol
IAM Identity and Access Management AWS permissions
IaC Infrastructure as Code Terraform
JWT JSON Web Token Authentication token
LLM Large Language Model Claude Sonnet 4.5
MAU Monthly Active Users Cognito pricing
OIDC OpenID Connect Authentication protocol
PITR Point-in-Time Recovery DynamoDB backups
RAG Retrieval-Augmented Generation AI architecture pattern
S3 Simple Storage Service Object storage
SLA Service Level Agreement Uptime guarantee
SNS Simple Notification Service AWS alerts
SSE Server-Side Encryption Encryption at rest
TLS Transport Layer Security Encryption in transit
VPC Virtual Private Cloud AWS network isolation

Last updated: 2026-01-01