Vector embeddings are just numbers. They put data like text or images into a multi-dimensional space so AI agents can handle unstructured information.
### 1. What are Vector Embeddings?
**Vector embeddings** map data to coordinates in a high-dimensional space. Agents use them to give semantic weight to documents, chat logs, or knowledge bases.
* **Semantic Similarity:** Things with similar meanings sit close together. "Artificial intelligence" and "machine learning" are neighbors. "Artificial intelligence" and "banana" are far apart. * **Dimensionality:** Vectors usually have hundreds or thousands of dimensions, such as 768, 1536, or 3072. Each dimension tracks a latent feature the model learned during training.
### 2. How Do They Work for AI Agents?
LLMs have fixed context windows and no innate knowledge of private or real-time data. Embeddings bridge that gap.
#### Step 1: Ingestion and Embedding
When an agent needs external info, like an internal wiki or a product catalog, an **embedding model** processes it first.
* **Action:** The model turns raw text into a vector. * **Storage:** The vector goes into a **Vector Database** like Pinecone, Weaviate, Milvus, or pgvector.
#### Step 2: Query Transformation
The agent doesn't search by keywords when you ask, "What is our refund policy?"
* **Action:** It sends your query to the same embedding model to create a query vector.
#### Step 3: Similarity Search (Retrieval)
The agent compares that query vector against the stored vectors.
* **Metric:** It uses math like **Cosine Similarity** or **Euclidean Distance** to find the closest matches. * **Result:** The system pulls the most semantically relevant data chunks, like specific paragraphs from a policy document, that match your intent.
#### Step 4: Context Augmentation (RAG)
The agent injects those chunks into its prompt as **context**.
* **Action:** The LLM gets your question plus the retrieved info. * **Output:** The LLM answers based on that context. This grounds the response in real data instead of hallucinating from pre-training.
### Summary for Liminality Context
For your **Liminality** project, vector embeddings ground your routing MCP:
1. **Understand Intent:** It converts questions into vectors to match predefined sub-questions or tool descriptions. 2. **Retrieve Context:** It fetches historical data or docs to inform the LLM. This ensures question decomposition relies on facts, not generic assumptions.
This lets agents act as knowledgeable assistants with access to live information.
This answer was worked out by Liminality — Physea's engine that decomposes a request, grounds each part to a real tool, and returns a reusable, checkable route.
Connect it over MCP: https://mcp.physea.ai/mcp · physea.ai