RAGO Pipelines

To access this menu, go to the sidebar and select Pipeline & Preset > RAGO Pipelines.

You can manage all RAG (Retrieval-Augmented Generation) pipelines from this tab. This includes selecting a pipeline type, uploading a new pipeline, adding or editing presets, and configuring settings to control how your application retrieves and processes information.

---

Select Pipeline

Select the pipeline you want to use from the available options on the RAG Pipelines page. Each pipeline type offers different capabilities and is suited for specific use cases.

Pipeline Descriptions:

1. No-Op A basic application without a connected knowledge base. Ideal for simple conversations, testing, or as a default fallback pipeline.

2. Standard An application connected to a knowledge base that retrieves relevant information and generates responses based on the retrieved content. This Standard RAG pipeline focuses on retrieving semantically similar document chunks to support contextual answers.

3. Graph RAG An extended version of the Standard RAG pipeline that enhances traditional vector-based retrieval with knowledge graph integration. Graph RAG combines both vector-based chunks and structured graph entities or relationships extracted from documents to produce richer, more context-aware responses.

Available options may vary and could differ from the example shown below.

---

Upload Pipeline (Deprecated)

You can upload pipeline by clicking the "Upload Pipeline" button on the top right.

Here is a minimal example of the .zip file

327KB

custom-pipeline-example.zip

archive

Open

Add Preset

After selecting a pipeline, click New Preset to create a preset configuration.

Fill in the required fields below, then click Save to store your changes or Cancel to discard them.

1. General Information

   1. Preset ID Enter a unique identifier used to register and reference this preset within the system.

2. Model Configuration

   1. Supported Model IDs Select which LLMs (Large Language Models) this preset supports (e.g., GPT-4, Claude 3).

   2. Use Cache Choose whether to enable caching for faster repeated retrievals.

      1. Yes – Cache enabled.

      2. No – Cache disabled.

3. Privacy & Data Protection

   1. Anonymize Em Control whether PII masking is applied during embedding/retrieval.

      1. Yes – PII will be masked.

      2. No – No masking applied.

   2. Anonymize Lm Control whether PII masking is applied during generation (when prompts are sent to the LLM).

      1. Yes – PII will be masked.

      2. No – No masking applied.

   3. Support PII Anonymization Enable to hide or mask sensitive user data (PII) during the retrieval process.

      1. Yes – PII will be masked.

      2. No – No masking applied.

4. Retrieval & Knowledge Configuration

   1. Augment Context Enable to allow the application to pull information from a knowledge base before generating responses.

      1. Yes – Use knowledge base.

      2. No – Skip retrieval step.

   2. Chat History Limit Define how many previous messages the application retains in conversation memory.

   3. Prompt Context Char Threshold Set how much prior chat context is included when sending prompts to the model.

   4. Reference Formatter Threshold Set the minimum similarity score (range 0.0–1.0) required for a source to be cited in responses.

      1. Higher value = only highly relevant sources.

      2. Lower value = broader inclusion.

   5. Reference Formatter Batch Size Define how many candidate references are evaluated in each batch.

   6. Reference Formatter Type Choose the formatting style for displaying source references in generated answers.

   7. Use Model Knowledge Allow the model to use its built-in knowledge when no relevant information is found in the knowledge base.

      1. Yes – Fallback allowed.

      2. No – Must only use knowledge base.

5. Search & Retrieval Behavior

   1. Enable Smart Search Integration Integrate with the Smart Search Engine for enhanced semantic retrieval.

      1. Yes – Smart Search enabled.

      2. No – Smart Search disabled.

   2. Normal Search Top K Number of top results to retrieve using Standard RAG search.

   3. Rerank Kwargs Enter additional parameters (key-value pairs) to fine-tune reranking behavior.

   4. Rerank Type Choose the reranking method used to prioritize search results.

   5. Search Types Choose which search methods your application can use. You can select one or both.

      1. Normal – Standard RAG.

      2. Web – Live web search.

      3. SQL Search - Enable Chat Filter based on schema in DPO Pipeline.

      4. Essential Deep Research - Comprehensive research with a lite version that's faster and more concise.

      5. Comprehensive Deep Research - In-depth research with thorough analysis and explanations.

   6. Smart Search Top K Number of top results to retrieve using Smart Search.

   7. Vector Weight Set the weighting applied to retrieved vectors during ranking or scoring.

   8. Web Search Top K Number of top results to retrieve using Web Search.

6. Web Search Control

   1. Web Search Blacklist List of blocked domains that the application must avoid.

   2. Web Search Whitelist List of approved domains the application can access during web searches.

7. Document Processing

   1. Support Multimodal Enable if the preset should support both text and image input.

      1. Yes – Accepts text + image.

      2. No – Text only.

   2. Use DocProc Enable to process uploaded documents using the Document Processor (DocProc). This extracts and structures document content and supports file attachments for knowledge base ingestion.

      1. Yes – Document processing active and attachments supported.

      2. No – Document processing disabled.

8. Swirl Configuration

   1. Autosuggest Providers Provide real-time search suggestions as users type. Example:

      1. Enabled: { "[Unique Provider ID]": internal/web }

      2. Disabled: {}

   2. Discovery Providers Specify providers responsible for generating suggested questions or follow-up queries.

   3. Shingle Providers Generate short overlapping word sequences (shingles) to improve query understanding. Example:

      1. Enabled: { "[Unique Provider ID]": internal/web }

      2. Disabled: {}

   4. Swirl Providers Primary data sources queried when a user performs a search via the Swirl interface. Format: { [Unique Provider ID]: internal/web }

      1. Internal – Local data sources (e.g., Elasticsearch, Chroma).

      2. Web – External sources (e.g., Firecrawl, Google).

   5. Web Swirl Providers Define external web-based providers specifically for web search queries.

9. Graph RAG Configuration

   1. Llama Index Graph RAG Embedding Model Specify the embedding model used for Graph RAG operations.

   2. Llama Index Graph RAG LLM Model Specify the language model used for reasoning in Graph RAG.

   3. Graph RAG Implementation Define or select the Graph RAG implementation method used in your environment.

10. Memory Settings

    1. Enable Memory Allow the pipeline to recall past interactions for context.

    2. Retrieve Memory Threshold Minimum similarity score for retrieving past memory.

    3. Retrieve Memory Top K Number of past memory entries to retrieve.

11. Safety & Guardrails

    1. Allowed Topics List of topics the model is allowed to discuss.

    2. Banned Phrase Words or phrases the model must avoid.

    3. Core Restriction Categories

    4. Enable Guardrails Activate safety filters to prevent unsafe or restricted outputs.

    5. Guardrail Fallback Message Message shown if a response is blocked by guardrails.

    6. Guardrail Mode Choose where guardrails are applied: Input, Output, Both, or Disabled.

    7. Topic Safety Mode Enable topic-based safety checks during conversations.

12. Others

    1. Enable Live Chat

The available fields in the Add Preset form may vary depending on the selected pipeline type.

View and Edit Preset

To view or edit a preset, click the Edit icon on the Preset List page.

You can review and modify the preset details as needed. Once you’re done, click Save to apply the changes or Cancel to discard them.