ToolUp.RAG

Retrieval-augmented generation runtime for apps built on ToolUp Platform. Vector store, retrieval pipeline, background ingestion + reembedding services, prompt builder for injecting retrieved chunks into AI system prompts.

Wraps ToolUp.AI — RAGServerApp is a flat superset of AIServerApp, which is itself a flat superset of ServerApp. Pick the tier that matches your needs.

When to use this companion

• AI assistant should ground its answers in user-uploaded documents, notes, narrative-committed content, or other text data.
• "What does our team's data say about X?" — answering questions with citations from the team's corpus.
• Semantic search over user-tenant data without operator access to the content.
• Continuous re-indexing as new documents arrive.

When NOT to use this companion

• Exact-match keyword search — use SQL LIKE, full-text search (Postgres tsvector, etc.), or a dedicated search engine. RAG is for semantic relevance, not exact match.
• Cross-tenant search — the SDK enforces per-scope isolation. Building a global search across all tenants requires an IRetrievalPipeline impl that intentionally widens scope access, defeating the security property.
• High-throughput log analytics — vector stores aren't optimised for billion-row time-series. Use a proper log store (Loki, ClickHouse, Splunk).
• Single-shot embedding without retrieval — depend on IEmbeddingProvider from ToolUp.Platform.Server directly.

What's in the box

Two packages:

Plus embedding-provider sub-companions:

• ToolUp.EmbeddingProviders.Local — in-process TF-IDF; dev / CI / offline.
• ToolUp.EmbeddingProviders.OpenAI — text-embedding-3-small against OpenAI API.

And vector-store sub-companions:

• ToolUp.VectorStores.Hnsw — HNSW approximate-nearest-neighbour; lifts the ~50K-chunk ceiling of InMemoryVectorStore.

Quick start

Add the packages:

<PackageReference Include="ToolUp.RAG.Server" />
<PackageReference Include="ToolUp.EmbeddingProviders.OpenAI" />
<PackageReference Include="ToolUp.AIProviders.Claude" />

Wire the server composition root:

open ToolUp.RAG
open ToolUp.EmbeddingProviders.OpenAI

let embedder = OpenAIEmbeddingProvider.create secretStore :> IEmbeddingProvider

RAGServerApp.create (aiProviderFactory, aiConfigStore, embedder)
|> RAGServerApp.withConfig serverConfig
|> RAGServerApp.withAuth authProvider
|> RAGServerApp.addModules modules
|> RAGServerApp.withAITools AITools.allTools
|> RAGServerApp.run

The agent loop now augments its system prompt with retrieved chunks from any registered scope. Documents uploaded via KnowledgeBase (or any VectorisationHandler-registered DataType) get auto-indexed on the post-save hook.

See getting-started.md for the full walkthrough.

Concepts

See concepts.md for the vector store, retrieval pipeline, chunking, ingestion + reembedding services, scope isolation, prompt-builder composition.

API reference

See api-reference.md for RAGServerApp, IRetrievalPipeline, IVectorStore, IEmbeddingProvider, VectorisationHandler, RAGPromptBuilder, and the per-deployment tuning knobs (withTopK, withMinScore, withMergeStrategy, etc.).

Extending

See extending.md for writing a new embedding provider, vector store, or retrieval tracer.

Scope isolation

The retrieval pipeline filters requested scopes against the caller's AccessContext.TeamId before any call into IVectorStore. A mismatched Team teamId returns an empty result rather than an error — prompt builders compose without fault handling. Platform and Deployment scopes are universally readable for authenticated callers (when PlatformKnowledgeBase is enabled).

There's no API path to retrieve across team boundaries. Custom IRetrievalPipeline impls that bypass scope filtering should not exist — the safe default is structural.