Semantic
Bring meaning into your database with native LLM integration, embeddings, and semantic operations as first-class primitives.
Semantic is one of Tacnode’s three foundational pillars—alongside Shared and Live. While traditional databases store and retrieve data by exact values and relationships, modern AI applications need to work with meaning: similarity, intent, classification, and generation.
Tacnode makes semantic operations a native database capability through the Semantic Operator—an llm extension that lets you invoke LLM providers like OpenAI or Amazon Bedrock Claude directly from SQL. This ensures:
- Transactional consistency — Semantic operations participate in the same ACID transactions as your data mutations
- Unified interpretation — All consumers of the database see identical semantic results, eliminating drift and sync issues
- Simplified architecture — No external service orchestration, no result caching, no round-trip overhead
Semantic Operator vs. Vector Search
Tacnode supports both the Semantic Operator and vector search, but they serve different purposes:
| Vector Search | Semantic Operator | |
|---|---|---|
| What it does | Finds similar items by vector distance | Invokes LLM to reason about data |
| When computation happens | Embeddings pre-computed, search at query time | LLM called at query time |
| Use cases | Similarity search, recommendations, RAG retrieval | Classification, summarization, extraction, generation |
| Input | Vectors (numerical embeddings) | Text (natural language) |
| Output | Ranked results by distance | LLM-generated responses |
They work well together. Use vector search to retrieve relevant documents, then use the Semantic Operator to reason about or transform those documents.
Why Semantic Matters
When meaning lives inside your database:
- AI agents coordinate on shared state — Multiple agents read and write the same semantic representations with transactional guarantees
- Features stay consistent — Embeddings and classifications computed in the database are immediately available to all consumers
- Pipelines collapse — What used to require extract-transform-load becomes a single SQL statement