Retrieval memory for long-context work
Published Updated
This is the Learning Atlas from Orbit. Get your invite
AI providers and modelsHostinginfrastructuredatabases
-
Step 1: Embedding
Why this matters here
Embeddings come first because retrieval systems need a way to turn meaning into searchable vectors before they can recall anything useful.
In short
A numerical representation of text, images, or audio that captures the underlying semantic meaning of the data, allowing computers to understand relationships between concepts.
-
Step 2: RAG
Why this matters here
RAG uses those vectors to bring outside knowledge into a model session, making it the first complete pattern after embeddings.
In short
An architecture that improves AI responses by searching an external database for relevant facts and inserting them into the prompt before the model answers.
-
Step 3: Pinecone
Why this matters here
Pinecone gives the abstract retrieval pattern an external managed database where those vectors can actually live.
In short
A fully managed, serverless vector database purpose-built to store, search, and manage high-dimensional embeddings for AI applications.
-
Step 4: Chroma
Why this matters here
Chroma adds a local or developer-friendly alternative, helping learners compare managed retrieval with lighter-weight setups.
In short
An open-source, embedding database optimized specifically for developer experience and rapid integration into local AI applications.
-
Step 5: Context window
Why this matters here
Context window matters after retrieval because it explains the hard limit that retrieved material still has to fit inside.
In short
The maximum amount of text (measured in tokens) an AI model can process and remember in a single interaction.
-
Step 6: Token
Why this matters here
Tokens are the unit that makes the context limit measurable, so they close the loop from search result to model input budget.
In short
The fundamental unit of data processed by an LLM, roughly equivalent to a word or part of a word in English.