Build an intent recognition application flow using LangStudio & RAG

Background

The intent recognition customer service application flow uses intent recognition and RAG to automatically understand user needs and retrieve answers from a knowledge base. It applies to scenarios such as e-commerce, banking, and telecommunications, and developers can extend it flexibly.

Prerequisites

LangStudio supports Faiss and Milvus as vector databases. If you plan to use Milvus, you must create a Milvus instance. Create a Milvus Instance and Instance Management.

Note
Faiss works for testing without a separate database. For production, use Milvus to handle larger-scale data.
The RAG knowledge base corpus has been uploaded to OSS.

1. (Optional) Deploy LLM and embedding models

The application flow requires large language model (LLM) and embedding model services. If you already have a service that supports the OpenAI API, skip this step.

Go to Quick Start > Model Gallery and deploy models for the following two scenarios. For more information about deployment, see Model Deployment and Training.

For Scenario, select Large Language Models. For this example, deploy DeepSeek-R1. For more information about deployment, see One-click Deployment of DeepSeek-V4 and DeepSeek-R1 Models.

Note
In this example, the Intent Recognition, Question Rewrite, and Answer User Question nodes share the same model connection, so you only deploy DeepSeek-R1 once. In production, deploy separate models per node for better accuracy.
For Scenario, select Embedding. For this example, deploy the bge-m3 general-purpose vector model.

2. Create connections

2.1 Create an LLM service connection

This LLM service connection uses a model service (EAS service) deployed from QuickStart > Model Gallery. Other connection types are described in Connection Configuration.

In LangStudio, select a workspace. Under Configure Service Connection, go to the Model Service tab and click New Connection to create a generic LLM service connection.

Key parameters:

Parameter	Description
Model name	You can obtain the model name from the model details page. To open this page, click the model card on the Model Gallery page. For more information, see Create a Connection - Model Service.
Service provider	PAI-EAS Model Service: Used in this example. For EAS Service, select the LLM service deployed in Section 1. The `base_url` and `api_key` auto-populate with the VPC access address and token of the deployed LLM. Third-party Model Service: Use a third-party model service. For example, if using the official DeepSeek service, the `base_url` is `https://api.deepseek.com`, and the `api_key` can be obtained from the DeepSeek official website.

2.2 Create an embedding model service connection

Similar to Section 2.1, create a generic embedding model service connection.

Set the Service Provider to PAI-EAS Model Service, and configure the EAS Service, Model Name, base_url, and api_key fields.

3. Create a knowledge base index

Creating a knowledge base index involves parsing, chunking, and vectorizing the corpus, then storing the results in a vector database. Key parameters are described below. For more information about other configurations, see Knowledge Base Management.

Parameter	Description
Basic configuration
Data source OSS path	Specify the OSS path to the RAG knowledge base corpus prepared in the Prerequisites section.
Output OSS path	Path for intermediate results and index data generated during parsing. Important With Faiss, index files are saved to OSS. If you use the default PAI role (the Instance RAM Role set when starting the runtime for application flow development), the flow can access your workspace's default storage bucket. Set this parameter to a directory within the OSS bucket that contains your workspace default storage path. If you use a custom role, grant it OSS access (the AliyunOSSFullAccess policy is recommended). Manage RAM Role Permissions.
Embedding model and database
Embedding type	Select General Embedding Model.
Embedding connection	Select the embedding model service connection created in Section 2.2.
Vector database type	Select Faiss. This document uses a Faiss vector database as an example.

4. Create and run the application flow

Go to LangStudio, select a workspace, go to the Applications tab, and click Create Application to create an Intent Recognition Customer Service application flow.

On the New Application Flow page, select the Create from Template tab, enter a Name (for example, intent-002), and select a runtime.
Start the runtime: Click Create Runtime in the upper-right corner and configure it. Note: The runtime must be running before you can parse Python nodes or view more tools.

Configure parameters including Runtime name, Resource specification, Security group, vSwitch, Mount configuration, and Maximum duration per run. Select a VPC from the VPC configuration dropdown.

VPC configuration: If you chose Milvus as the vector database type in Step 3, select the same VPC as the Milvus instance or ensure VPC interconnection. Faiss does not require VPC configuration. This example uses Faiss.
Develop the application flow.

Keep defaults for other settings, or adjust as needed. Key node configurations:
- Intent Recognition: Uses an LLM to analyze user input and route to the corresponding branch.
  - Multi-intent configuration: Define intents with clear, non-overlapping descriptions. This example uses the default intent "Questions about Alibaba Cloud PAI products".
  - Model Settings: Select the connection created in Section 2.1.
  - Chat History: Toggle to use conversation history as an input variable.
- Question Rewrite: Uses an LLM to rewrite the matched intent information.
  - Model Settings: Select the connection created in Section 2.1.
  - Chat History: Use the same configuration as the Intent Recognition node.
- Knowledge Base Retrieval: Retrieves text relevant to the user's question from the knowledge base. For more information, see Knowledge Base Management.
  - Knowledge base index name: Select the knowledge base index created in Step 3.
  - Top K: Number of top matching entries to return.
- Answer User Question: Uses an LLM to answer the user's question.
  - Model Settings: Select the connection created in Section 2.1.
  - Chat History: Use the same configuration as the Intent Recognition node.
- Other Intent: Runs custom Python code for complex data processing.
- Variable Aggregation: Merges outputs from different intent branches into a unified variable.
For more information about each node component, see Pre-built Component Reference.
Debug/Run: Click Run in the upper-right corner to run the application flow. For common runtime issues, see FAQ.

A chat test panel appears on the right. Enter test questions to verify the behavior of the application flow. For example, entering "hello" returns "Sorry, I can only answer questions about Alibaba Cloud PAI." Entering "What are the features of Alibaba Cloud PAI products?" returns a feature list including AI computing resource management, AI workspace, and AI asset management (dataset management, model management, task management), confirming the flow works correctly.
View Trace: Click View Traces below the generated answer to see the trace details or topology view.

The Trace details page shows a waterfall chart of span durations — for example, intent-002 (CHAIN/17.41s), intent_recognition (TASK/717.25ms), ChatCompletion (LLM/707.14ms), rewrite (TASK/369.96ms), and index_lookup (TASK/429.68ms). Click a span to view its details, including application name, API name, spanId, status code, and input/output data. The top of the page shows the TraceID, start time, total duration, and total tokens.

5. Deploy the application flow

On the application flow development page, click Deploy in the upper-right corner to deploy the flow as an EAS service. Keep the default values for other settings or adjust them as needed. Key parameters:

Resource deployment > Number of instances: Specify the number of service instances. The deployment in this topic is for testing only, so set this parameter to 1. In a production environment, configure multiple service instances to reduce the risk of a single point of failure.
VPC > VPC: If you chose Milvus in Step 3, configure the same VPC as the Milvus instance or ensure VPC interconnection. Faiss does not require VPC configuration. This example uses Faiss.

For more information about deployment, see Deploy an Application Flow.

6. Call the service

After a successful deployment, you are redirected to PAI-EAS. On the Debug tab, configure and send a request. The request body key must match the "Dialogue Input" parameter in the flow's "Start" node. This example uses the default field, question.

For example, enter {"question":"hello"} in the request body and click Send Request. A 200 status code and the response {"output":"Sorry, I can only answer questions about Alibaba Cloud PAI."} confirm successful debugging.

In raw mode, enter a JSON object such as {"question":"What are the functions of Alibaba Cloud PAI DSW"}. A Status Code: 200 and JSON response with the answer indicate a successful service call.

Other calling methods, such as API calls, are described in Deploy an Application Flow - Call the Service.