FlowRag — RAG Knowledge Node

What is Retrieval-Augmented Generation (RAG)?

RAG is an AI pattern that enhances language model responses by first retrieving semantically relevant documents from a knowledge store, then including those documents as context in the AI prompt. This allows AI nodes to answer questions grounded in your specific, up-to-date business knowledge — rather than relying solely on model training data.

FlowRag handles the knowledge storage and management half of this pattern. It converts text into vector embeddings, stores them in a configured vector database, and keeps that knowledge base current through insert, update, and delete operations.

What this node does: FlowRag provides managed knowledge base operations within BizFirst workflows. Insert knowledge documents (policies, FAQs, product descriptions, contract clauses) into named collections, keep them current with update and delete operations, and expose the collection to AI agents for semantic retrieval at query time. Supports Pinecone, Weaviate, Chroma, Qdrant, and pgvector as vector store providers.

Supported Operations

knowledge/insert

Add a new document to a knowledge collection. The text is embedded and stored with optional filename and source metadata. Returns the assigned knowledgeId and chunk count.

knowledge/delete

Remove a document from a collection by its knowledgeId GUID. The embedding vectors and all associated chunks are permanently deleted from the vector store.

knowledge/update

Replace the content of an existing document. The old embedding is removed, the new content is re-embedded and stored. Returns the old and new knowledgeId values.

Connection Configuration

These fields are shared across all three operations and identify the vector store and collection to operate on.

Field	Required	Description
`providerName`	Required	Vector store provider key. Supported values: `pinecone`, `weaviate`, `chroma`, `qdrant`, `pgvector`.
`collectionName`	Required	The collection or index name within the vector store where documents are stored. All items in a collection must use the same embedding model.
`credentialId`	Optional	Credential ID for the vector store API key or connection string, stored in the BizFirst Credentials Manager.
`embeddingProvider`	Optional	LLM provider used to generate embeddings, e.g. `openai`. Defaults to the workspace-level embedding configuration.
`embeddingCredentialId`	Optional	Credential ID for the embedding model API key, if different from the vector store credential.
`embeddingModel`	Optional	Embedding model identifier, e.g. `text-embedding-3-small`, `text-embedding-3-large`. Must be consistent across all items in a collection.

Vector Store Providers

Provider	Config Value	Best For	Considerations
Pinecone	`pinecone`	Managed cloud vector DB, production RAG at scale	Fully managed; requires Pinecone account and API key
Weaviate	`weaviate`	Schema-rich vector store with built-in modules	Self-hosted or cloud; strong metadata filtering
Chroma	`chroma`	Lightweight, developer-friendly, local or cloud	Good for prototyping; scales less than Qdrant/Pinecone
Qdrant	`qdrant`	High-performance, rich payload filtering	Self-hosted or Qdrant Cloud; REST + gRPC APIs
PostgreSQL pgvector	`pgvector`	Teams already running PostgreSQL	Requires pgvector extension; performance degrades beyond ~10M vectors

When to Use

Company policy Q&A bot: Insert HR, IT, and operations policies into FlowRag at publish time. When an employee asks a question, a downstream AI node retrieves the relevant policy sections as context.
Customer support FAQ ingestion: Use knowledge/insert in a workflow triggered whenever a new FAQ article is published in your CMS. Keep the knowledge base current automatically.
Document library maintenance: When a policy document is revised, call knowledge/update to replace the old embedding with fresh content — no re-index of the entire collection required.
Expired content cleanup: When a product is discontinued or a policy is retired, call knowledge/delete to prevent the AI from retrieving outdated information.
Bulk initial ingestion: Use a Loop node to iterate over a document array and call FlowRag knowledge/insert on each item to seed a new knowledge collection.

Embedding Model Selection

Model	Provider	Dimensions	Best For
`text-embedding-3-small`	OpenAI	1536	Default — excellent balance of quality and cost for general business content
`text-embedding-3-large`	OpenAI	3072	Highest accuracy for technical, legal, or scientific content where precision matters
`text-embedding-ada-002`	OpenAI (legacy)	1536	Legacy compatibility only — prefer `text-embedding-3-small` for new collections

Embedding model consistency: All documents in a collection must use the same embedding model. Mixing models produces invalid similarity scores at query time. To change the model, delete all items and re-insert with the new model.

Knowledge Base Design Tips

Split long documents into focused, single-topic chunks of 200–500 tokens before inserting — retrieval quality improves significantly with tightly scoped chunks
Set the source field to the source URL or file path so downstream AI nodes can cite their sources in responses
Use the fileName field consistently to enable targeted bulk operations (e.g. delete and re-insert all chunks from a specific file when the file changes)
Maintain separate collections per domain (HR, products, legal, support) for better retrieval precision and to avoid cross-domain noise
Store the returned knowledgeId in your CMS or database record so you can reliably reference it for future update and delete operations

Combine with FlowAiAgent for Autonomous RAG

Register a FlowRag query as a tool in the BizFirst Tool Registry, then give it to a FlowAiAgent node. The agent will autonomously decide when to query the knowledge base — calling it multiple times with different queries if the first retrieval does not fully answer the goal. This enables knowledge-augmented agents that can both reason and retrieve without requiring you to hard-code which knowledge base to query at each step.

Common RAG Pipeline Pattern

Content trigger — CMS publishes or updates a document
FlowRag knowledge/insert — embed the document and store in the appropriate collection
Store knowledgeId — save the returned GUID to the CMS record for future update/delete
Query time — user submits a question via a form or chat trigger
FlowAiAgent or AI Chat — retrieves relevant chunks from FlowRag (via tool call), uses context to generate a grounded response
Response — returns the AI answer with citations from the source metadata field