RAG Quickstart : custom settings

Create a RAG application with full control over ingestion, chunking, vectorization, and retrieval settings.

Custom settings give you control over every stage of the RAG pipeline — from data ingestion and chunking strategy to vector storage and retrieval configuration. Use this path for production workloads, specialized document types, or when default settings do not meet your accuracy requirements.

From the RAG Applications dashboard, click Create. The dialog asks whether to use default or custom settings.

Select Create with custom settings. The wizard walks you through four steps: LLM → Ingest → Chunk & Vectorize → Retrieval.

---

Step 1: Configure LLM

Enter a name for your RAG application in the RAG Application Name field. Names must use only underscores (_) and alphanumeric characters.

Enter a name for your RAG application in the RAG Application Name field. Names must use only underscores (_) and alphanumeric characters.

In the LLM table, select the model your application will use for answer generation. The table lists all LLMs configured in your workspace, including their type, model name, deployment type, and status. Mark one as default if needed.

Step 1: Enter an app name and select an LLM from your configured models

Click Next to proceed to ingestion.

---

Step 2: Ingest

Choose how to get your data into the application.

Option A — Ingest new data: Upload new documents or connect to a data source for the first time.

Step 2: Choose to use an existing dataset already in the system, or ingest new data

If you select Ingest new dataset, choose a data source type:

• Pick from Data Storage — pull from a connected storage (S3, GCS, SharePoint, etc.)

• Links and Web Crawling — crawl public or authenticated web pages

• Upload from Computer — upload files directly (up to 5 MB each)

Select a data source type for new ingestion

After selecting a source type, configure the File format, an optional Parse Setting for information extraction, and the Data Storage destination.

Configure file format, parse setting, and data storage for the ingestion

Choosing a data source: The Ingest step first asks how you want to get started — use a dataset already in the system, or ingest a new one.

Choose to use an existing ingested dataset or ingest new data from a connected source

If you select Use existing dataset, the Dataset Explorer opens so you can browse and select any previously ingested dataset across all workspaces.

Browse all available datasets across workspaces and select one to use as the data source

If you select Ingest new dataset, configure the File format. The dropdown lists all supported types: Audio/Video, Image, Excel, Markdown, PDF, Plain Text, and PowerPoint.

Select the file format — choose one or more types that match your source documents

After selecting a file format, the Data Storage field appears. Select a connected storage connector from the dropdown.

File format selected — now choose the data storage connector for your source files

Once you select a connector, the folder structure for that storage expands beneath the Data Storage field.

The connected storage expands to show its folder structure — browse to your target directory

Select the specific files to index. Use the File pattern field to filter by filename. Check each file you want to include, then optionally enable Automatically add new files to keep the app updated as new documents arrive.

Select the files to ingest — use file pattern filtering and enable auto-update for ongoing ingestion

After configuring the data source, the Ingest step shows a summary of the Source and Target configuration cards. Verify the source file type, storage connector, and target dataset location before proceeding.

Review the source and target configuration — confirm file type, connector, and target storage before advancing

To change the target dataset location, click the edit icon on the Target card. The Target Configuration panel opens with a directory browser.

Select the target ingestion directory from the directory browser — choose an existing location or create a new one

Click Next to proceed to chunking and vectorization.

---

Step 3: Chunk & Vectorize

Select the dataflow that defines how documents are split into chunks and converted into vector embeddings. Each card shows the chunking strategy, chunk size, overlap, embedding model, and target vector dataset.

Step 3: Select a pre-configured chunking and embedding dataflow, or configure a custom one

Each card displays the chunking strategy, chunk size, overlap, embedding model, and target dataset. Select a card and click Next to proceed with those settings, or click Choose Target Dataset to assign a different vector storage target.

Select a dataflow card — use the Choose Target Dataset button to assign a different vector storage target

Clicking Choose Target Dataset opens a modal. Select the Vector database type, then choose the Workspace and Store in Collection (collection name).

The Choose Target Dataset modal — select a vector database and collection to store embeddings

Use the dataset browser to find and select the target dataset. The tree shows all available datasets in the selected collection.

Browse and select the target dataset from the list — click Save to confirm

Click Next to proceed to retrieval configuration.

---

Step 4: Retrieval

Configure the retrieval pipeline — how your app finds relevant chunks and generates answers. The pipeline diagram at the top shows the full flow: Source → Embeddings → Encoder & Chunking → Query Retriever → Generation.

Step 4: The retrieval configuration panel with the pipeline diagram and sub-sections in the left menu

Source

Configure which dataset columns map to the question, answer, external link, and metadata fields. Enable Auto Extraction of Filters from query to let the system extract filter values automatically at query time.

Map dataset columns to question, answer, external link, and metadata fields

After configuring the source columns, click Modify Metadata Fields to select which metadata fields are available for filtering at query time.

Select the metadata fields to expose as query-time filters — for example, company_name and fiscal_year

The source panel also lets you enable Auto Extraction of Filters from query. When enabled, the system reads the query text and automatically applies matching metadata filters without requiring the user to set them manually.

Enable auto extraction of filters so the system applies metadata filters automatically from the query text

Configure which metadata fields the system uses for automatic filter extraction

Embeddings

Select the embedding model that converts document chunks into vector representations for semantic search.

Select the embedding model — the default is all-mpnet-base-v2, a general-purpose sentence embedding model

Encoder & Chunking

Configure the vector database and chunking storage settings. Specify the vector database type, storage accounts, collection names, and index names. These settings define where chunked and embedded data is stored.

Configure the vector database, storage accounts, MongoDB collection, and index name for chunk storage

Input Guardrails

Select an input guardrail to apply before queries reach the retrieval pipeline. Guardrails can filter or block queries that violate safety rules or topic restrictions.

Optionally assign an input guardrail to filter or block queries before they reach the LLM

Query Rewriting

Enable query rewriting to improve retrieval recall. When active, the system reformulates or expands the user's query before searching the vector store. Select the LLM to use for rewriting.

Select an LLM to power query rewriting — the system rewrites queries to improve retrieval accuracy

Query Retriever

Configure the hybrid search strategy, similarity metric, similarity threshold, chunk retrieval type, and optional re-ranking. These settings determine how candidate chunks are scored and ranked before being passed to the LLM.

Set the search type (for example, Hybrid Search and Merge), similarity metric, and custom concepts for retrieval

Set the similarity threshold, chunk retrieval type, MMR diversification toggle, and LLM-based re-ranking

Caching & External Memory

Enable LLM response caching to avoid re-running the retrieval pipeline for identical queries. Enable session tracking to maintain chat history context across turns.

Enable response caching and session tracking to improve performance and enable multi-turn conversations

Feedback/RLHF

Enable user-level feedback and domain expert feedback collection. Collected feedback supports reinforcement learning from human feedback (RLHF) workflows.

Enable user and expert feedback collection to support continuous improvement of your RAG application

Response Evaluation

Enable automatic evaluation of LLM-generated responses. When active, each response is scored for groundedness, context relevance, and answer relevance.

Enable response evaluation to automatically score every response for quality and factual grounding

Generation

Set the LLM provider and model used for answer generation. Provide a System Prompt to guide the model's behavior, and configure the error message shown when no matching context is found.

Configure the generation LLM, system prompt, and no-context error message

Review and Submit

When all retrieval settings are configured, click Finish update (or Review and Submit on new apps) to review a summary of the full pipeline configuration before saving.

Review the complete pipeline configuration — source data, chunking strategy, embedding model, and retrieval settings — before saving

Click Save to create your RAG application. The app appears in the dashboard with status indicators for ingestion, vectorization, and retrieval.