Learned knowledge
Compressed into model parameters during training. Broad and fast to use, but difficult to update precisely or trace to a source.
Layer 5 — AI Infrastructure · Chapter 3
AI capability depends heavily on the information a system can learn from and access. Data infrastructure supplies that information in forms suitable for training, evaluation, retrieval, and real-time operation.
Raw data is rarely ready for AI. It may be duplicated, mislabeled, outdated, private, biased, corrupted, or legally restricted. Data infrastructure transforms it through a controlled pipeline.
Training changes the model’s internal parameters. Retrieval supplies information at the time of a request without retraining the model. This distinction matters because many facts change quickly, some data is private, and some knowledge must be cited or access-controlled.
Compressed into model parameters during training. Broad and fast to use, but difficult to update precisely or trace to a source.
Fetched from documents, databases, or services during use. Easier to update, cite, restrict, and inspect.
A retrieval system normally receives a question, searches for relevant material, ranks the results, and inserts selected evidence into the model’s context. The model then uses that evidence when producing an answer. Effective retrieval may combine keyword search, vector similarity, metadata filters, graph relationships, database queries, and reranking models.
An embedding represents an item as a numerical vector that captures aspects of meaning or similarity. A vector index allows the system to find nearby items efficiently. This is useful for semantic document search, recommendations, clustering, deduplication, and memory systems.
Vector similarity is not sufficient by itself. Production retrieval often needs:
Models can generate examples, labels, simulations, critiques, or edge cases. Synthetic data can reduce cost and expand coverage, but it may also reproduce the generating model’s errors, narrow the distribution, or make evaluation circular. Strong pipelines mix generation with filtering, diversity checks, expert review, and validation against real-world outcomes.
Governance establishes who may use data, for what purpose, for how long, and with what safeguards. It includes consent, retention, deletion, geographic restrictions, contractual rights, lineage, audit trails, and procedures for correcting or removing problematic data.
In many domain-model businesses, the durable advantage is not the model alone. It is a lawful, high-quality, continuously improving data system.