Configure Graph RAG

Graph RAG is a retrieval tool you add to a Touchpoint. It searches your Knowledge for relevant resources, then follows the relationships between them to find more. This guide takes you from a new Knowledge to a working, tuned Graph RAG tool.

📖WHAT IS GRAPH RAG?

A basic (naive) RAG search finds resources that match the meaning of a query. Graph RAG does that first, then traverses the knowledge graph to pull in related resources the meaning-only search would miss.

Prerequisites

Before you start, make sure you have:

• Access to the AI Control Room Settings area.
• An embedding model available in the AI Model Registry.
• Content (documents or other resources) ready to add to a Knowledge.

How the setup works

Graph RAG comes together in three parts:

1. Create a Knowledge: the set of resources to search, plus the rules for how resources are processed.
2. Configure its Touchpoint: the model and instructions for answering questions.
3. Tune the Graph RAG tool: how broad and how strict the retrieval is. You do the first two once. You return to the third whenever you want to adjust quality.

Create a Knowledge

A Knowledge holds your resources and defines how they are processed when added. Creating one is the first step, because Graph RAG searches a Knowledge.

1. Go to Settings → Knowledge and select + New.
2. Give the Knowledge a name and an optional description.
3. Open the AI Search tab to set up processing.

Set up AI Search

The AI Search tab controls how resources are prepared for retrieval. Configure these in order:

1. Turn on Allow AI search. This enables the RAG process for the Knowledge.
2. Select an Embedding model. This model converts text into vectors so the system can measure how similar a resource is to a query.
3. Choose a Chunking type. Chunking splits each resource into smaller pieces before embedding. See the chunking reference below.

📝PROCESSING APPLIES TO TEXT

Chunking, embedding, and keyword extraction currently run on textual resources only.

Set up keyword extraction (optional)

Keyword extraction tags each chunk with its most representative terms. Graph RAG uses these terms as an extra signal when ranking results.

1. Turn on Allow Keywords Extraction.
2. Set the maximum number of entities per chunk.
3. Choose an extraction approach:

Approach	How it works	When to use
Rule-based	Counts the most frequent terms after cleaning and lemmatizing the text.	Faster, lower precision.
LLM-based	A language model reads each chunk and picks the key terms. Uses the Touchpoint's model.	Better quality, slightly slower.

With the LLM-based approach you can edit the system prompt sent to the model, which helps when terms should be domain-specific.

Configure the Touchpoint

Creating a Knowledge also creates a Touchpoint with default settings, including a working Graph RAG tool. Adjust it in Settings → Touchpoints.

• Conversational Model: the model that writes the answers. Choose from the available models.
• System prompt: the instructions for that model. You can edit the default. By default, the model answers in a set order: it checks the structure of the data, runs a SQL query if the answer is in a database, and falls back to the Graph RAG tool for anything else.

Tune the Graph RAG tool

The Graph RAG tool runs in three phases. Each phase has settings that trade breadth against precision.

Core settings

• Knowledge IDs: the Knowledge bases this tool searches. Select one or more.
• Use Graph Expansion: the master switch for following relationships. Off means a basic semantic search only.

Phase 1: semantic search

The tool first finds resources that match the meaning of the query.

Setting	What it does	Effect
Initial Search Limit	How many results to fetch before filtering.	Higher (e.g. 200) widens the pool but adds latency.
Semantic Threshold (0.0 to 1.0)	Minimum similarity score to keep a result.	Lower (0.3) = more results; higher (0.7) = more precise.

Phase 2: graph expansion

The tool takes the strongest Phase 1 results and follows their relationships to find related resources.

Setting	What it does	Effect
Graph Expansion Threshold (0.0 to 1.0)	Minimum Phase 1 score for a resource to have its relationships explored.	Keeps low-relevance results from pulling in irrelevant neighbors.
Relationship Types	Which kinds of relationships to follow (for example "is part of", "references").	Leave empty to follow all types.
Max Graph Results (minimum 1)	How many related resources to retrieve per starting resource.	Controls breadth and performance.

Phase 3: final ranking

The tool combines results from both phases, scores them, and fits them into the model's context window. Phase 2 resources get a hybrid score that blends meaning and keyword overlap.

The hybrid score is calculated as:

combined_score = (semanticWeight * semantic_score) + (keywordWeight * keyword_overlap_score)

Setting	What it does	Effect
Semantic Weight / Keyword Weight (must add to 1.0)	Balances meaning against exact-term overlap.	High Semantic Weight (0.8) favors topic; high Keyword Weight (0.8) favors matching terms.
Keyword Overlap Method	The algorithm for scoring term overlap.	Fuzzy is recommended (handles partial matches and typos). Jaccard, dice, and cosine are also available.
Relevance Threshold (0.0 to 1.0)	Final filter for Phase 2 resources.	Only resources at or above this score are kept.
Target Context Utilization (0.5 to 0.95)	Share of the context window to fill with retrieved text.	Leave room for the prompts and the model's answer (e.g. 0.9 fills 90%).

Result

Your Touchpoint can now answer questions using Graph RAG: it finds resources by meaning, expands the search across relationships, and ranks the combined set before answering.

Chunking types reference

Chunking splits a resource into smaller pieces before embedding. Choose the type when you set up AI Search.

Type	How it splits	Notes
Semantic Chunker	Detects shifts in meaning using embeddings.	Best retrieval quality, but slower and needs an embedding model.
Sentence Chunker	Splits at sentence boundaries, within size limits.	Keeps sentences whole.
Token Chunker	Splits into fixed-length token windows.	Simple, size-based.
SDPM Chunker	Semantic chunking plus a second pass that merges related chunks.	Preserves context across gaps.