# AI Models

# AI Models

Curiosity Workspace uses AI models in three common ways:

• Understanding: NLP pipelines that extract structure from text (entities, signals, links).
• Retrieval: embeddings used for semantic similarity (vector search) and re-ranking.
• Generation: LLMs used for synthesis, assistance, and workflow automation.

The important architectural point: AI features are most reliable when they are grounded in your workspace data via graph + search retrieval.

# Embedding models (semantic similarity)

Embedding models map text (and sometimes other modalities) into vectors. In Curiosity Workspace, embeddings are used for:

• vector search (find semantically similar items)
• clustering and similarity (group related items)
• candidate generation for AI-assisted workflows

Design considerations:

• choose which fields get embeddings (usually long, descriptive text)
• enable chunking when fields exceed model context limits
• decide whether vector search is a primary retrieval method or a supplement to text search

See NLP → Embeddings and Search → Vector Search.

# NLP pipelines (extraction + enrichment)

NLP pipelines transform raw text into structured outputs, such as:

• extracted entities (people, products, IDs)
• normalized tokens and language-specific parsing
• optional entity linking into your graph

This enables:

• better filters (entities become facets)
• better graph navigation (mentions → resolved entities)
• better retrieval grounding for LLMs

See NLP → Overview and NLP → Entity Extraction.

# LLMs (generation + orchestration)

LLMs are typically used to:

• answer questions using retrieved context
• summarize, classify, or extract structured outputs
• drive multi-step workflows (tools, endpoints, actions)

Recommended patterns:

• retrieval first: fetch relevant nodes/documents before prompting
• tooling: move business logic into endpoints/tasks rather than relying on prompts alone
• auditability: store inputs/outputs where needed (policy dependent)

See AI & LLMs → Overview and AI & LLMs → Prompting Patterns.

# Safety and governance (conceptual)

Production AI typically needs:

• permission-aware retrieval
• logging/auditing for sensitive workflows
• strict separation of admin-only capabilities

See Administration → Security.

# Next steps

• Configure AI/LLM behavior: AI & LLMs → LLM Configuration
• Design semantic retrieval: Search → Hybrid Search