DocumentationRAG

Ground every answer in your knowledge.

Select a retrieval architecture, configure its models and search limits, then index the documents that should ground responses.

Configuration at a glance

6 retrieval patterns

Every RAG requires a name, embedding model, and vector dimension. Graph, router, vision, and synthesis models appear only when their capabilities are enabled.

Overview

Build a retrieval pipeline

A RAG configuration converts source documents into vectors and retrieves relevant context for each query. More advanced patterns add reranking, visual extraction, graph traversal, lexical search, or intent-based routing.

1. Architect

Choose the retrieval strategy for the shape of your knowledge.

2. Configure

Select models, vector dimensions, and bounded retrieval parameters.

3. Ground

Upload documents, wait for indexing, and inspect retrieval with real queries.

Step 01

Select a RAG pattern

Each pattern activates a different retrieval path and exposes only its relevant model and tuning controls in step two.

Naive RAG

Direct semantic embedding search followed by grounded response generation.

Retrieve-and-Rerank

Retrieves a broad candidate pool, then applies neural reranking to keep the strongest passages.

Multimodal RAG

Combines text retrieval with descriptions extracted from images, charts, and diagrams.

Graph RAG

Extracts entities and relationships, then traverses a knowledge graph for multi-hop context.

Hybrid RAG

Fuses dense vector similarity with sparse BM25 keyword matching for balanced recall.

Router RAG

Classifies each query and routes it between vector search, graph traversal, or direct response.

Step 02 · Section 01

RAG Details

Identify the knowledge base and describe the domain covered by its indexed sources.

RAG NameRequired

The display name used in RAG directories and selectors for agents, widgets, and API targets.

Use a domain-oriented name such as Enterprise Knowledge Base or Technical Manual Search.

DescriptionOptional

A concise summary of the knowledge domain, source material, and retrieval purpose.

This helps collaborators choose the correct knowledge base; it is not inserted into retrieved context.

Step 02 · Section 02

Vector Embedding Model

The embedding model converts document chunks and queries into numeric vectors. The selected dimension fixes the vector shape used for indexing and similarity search.

Embedding ModelRequired

Selects the provider model used to encode both document chunks and incoming queries.

Use the same embedding configuration throughout the lifetime of an index for comparable vectors.

Vector DimensionRequired

Sets the number of numeric values stored for each vector representation.

Available dimensions depend on the model; the current interface offers supported values such as 256, 512, 768, 1536, and 3072.

Step 02 · Section 03

Rerank AI Model

The rerank model applies a neural cross-encoder to score each candidate passage against the user query after broad vector retrieval. It re-orders the candidates so the final context keeps the most relevant passages.

Rerank ModelRequired

Selects the cross-encoder used to score vector-search candidates against each query.

Required for Retrieve-and-Rerank. It reorders the initial_top_k candidate pool before final_top_k passages are returned.

Step 02 · Sections 04–07

Specialized AI Models

Additional models appear only when the architecture or an optional capability needs them. Each compatible reasoning model can also expose an effort selector after selection.

Knowledge Graph Extraction Model section — Graph RAG · Entity and relation extraction

Query Intent Classifier Model section — Router RAG · Query classification

Enabled Vision and Media Description Model section — Optional · Visual asset extraction

Enabled Response Synthesis AI Model section — Optional · Grounded response synthesis

Graph ModelConditional

Extracts entity nodes, semantic relationships, and structured triples during indexing.

Required by Graph RAG and Router RAG.

Router ModelConditional

Classifies incoming intent and selects vector retrieval, graph traversal, or direct response.

Required only by Router RAG.

Vision ModelConditional

Describes images, charts, and diagrams so visual information can join the searchable context.

Optional toggle. Once enabled, selecting a compatible model becomes required.

Response ModelConditional

Turns retrieved passages into a grounded conversational answer for standalone RAG queries and widgets.

Optional toggle. When disabled, the pipeline returns formatted context passages directly.

Reasoning EffortConditional

Applies a supported effort level to graph, router, vision, or response models.

Shown only when the selected model publishes reasoning-effort options.

TemperatureConditional

Controls variation in generated output for applicable generative stages.

Default: 0.2. Range: 0.0–2.0; unavailable for models that do not support temperature.

Step 02 · Section 08

Architecture & Retrieval Parameters

Sliders control the amount of candidate context collected at each stage. The card changes with the selected pattern, so only parameters used by that pipeline are shown.

Router RAG Architecture and Retrieval Parameters section — Create RAG · Router retrieval parameters

Parameter	Pattern	Default / Range	Purpose
top_k	Naive, Multimodal, Router	5 · 1–50	Passages injected into the final context.
initial_top_k	Retrieve-and-Rerank	25 · 5–100	Broad semantic candidate pool before reranking.
final_top_k	Rerank, Hybrid	5 · 1–20	Passages retained after reranking or rank fusion.
vector_top_k	Graph	4 · 1–20	Semantic seed chunks used to locate graph entities.
graph_depth	Graph	1 · 1–3	Relationship hops traversed from the seed entities.
graph_limit	Graph	20 · 1–100	Maximum entity-relation triples added to context.
vector_k	Hybrid	15 · 1–50	Dense semantic candidates included in rank fusion.
bm25_k	Hybrid	15 · 1–50	Sparse lexical candidates included in rank fusion.

Step 02 · Section 08

Knowledge Base Documents

Save the RAG first to enable document uploads. Uploaded files are indexed with the selected embedding model and become available to retrieval after processing completes.

Changing the embedding model or vector dimension after indexing can require documents to be reprocessed so stored vectors remain compatible.

Step 02 · Section 09

Live Query Testing

After saving and indexing documents, submit representative queries to inspect the grounded answer, retrieved chunks, extracted facts, execution steps, and assembled context.

Required before save

Pattern, RAG name, embedding model, vector dimension, and every model required by the selected architecture.

After creation

Use questions that require exact facts and multi-source synthesis, then adjust retrieval limits based on the returned evidence.