Semarchy GenAI Gemini structured enricher

The Semarchy GenAI Gemini structured enricher extracts structured data from unstructured text to enhance data completeness and streamline the data entry process.

Plugin ID

Semarchy GenAI Gemini Structured Enricher - com.semarchy.engine.plugins.genai.gemini.structured

Description

The GenAI Gemini structured enricher is designed to extract structured data from unstructured text using Google Gemini language models. It can generate or extract up to 20 outputs in JSON format, including strings, booleans, numbers, and dates.

Prerequisites

To authenticate with the Vertex AI API, you need to have a Google Cloud service account and must configure xDM to integrate Gemini models.

  1. Create a Google Cloud service account.
    For detailed information, see the official Vertex AI documentation.

  2. At the end of the account creation process, download the service account key file.

  3. In xDM’s startup configuration, set the environment variable GOOGLE_APPLICATION_CREDENTIALS to the path of the service account key file that contains your credentials.

For additional information on how to get started with Vertex AI, see the LangChain4j documentation.

Plugin parameters

The following table lists the plugin parameters.

Parameter name Mandatory Type Description

Project

Yes

String

A globally unique, permanent identifier generated by the Google Cloud console. The project ID can be a combination of lowercase letters, numbers, and hyphens.

While it is possible to modify the project ID during project creation, we recommend using the one generated by the Google Cloud console.

Location

Yes

String

Geographical region where the Gemini model will be deployed and used. The following regions are supported:

  • Iowa (us-central1)

  • Las Vegas, Nevada (us-west4)

  • Montréal, Canada (northamerica-northeast1)

  • Northern Virginia (us-east4)

  • Oregon (us-west1)

  • Seoul, Korea (asia-northeast3)

  • Singapore (asia-southeast1)

  • Tokyo, Japan (asia-northeast1)

Model Name

Yes

String

Language model to be used.

To use the latest version, enter the model name without a version number, such as gemini-1.0-pro, gemini-1.5-pro, or gemini-1.5-flash.

Temperature

No

Number

Value ranging from 0.0 to 1.0 for balancing between conservative and coherent outputs (0) and creative variations (1) during text generation.

Max Output Tokens

No

Integer

Maximum number of tokens allowed in the generated output during text generation.
Default value: 50.

A token is about four characters, with 100 tokens equaling roughly 60-80 words. For shorter responses, specify a lower token value; for longer responses, specify a higher value. For reference, the Gemini Pro range is 1-8192 tokens (default: 8192), and the Gemini Pro Vision range is 1-2048 tokens (default: 2048).

Top K

No

Number

Value ranging from 1 to 40 for limiting the model’s predictions to the most probable tokens at each step of generation.

Top P

No

Number

Value ranging from 0.0 to 1.0 for defining the cumulative probability threshold for nucleus sampling (i.e., token selection).

Nucleus sampling is an alternative to temperature sampling. Ignore this parameter if you configured the Temperature parameter.

Max Retries

No

Integer

Maximum number of attempts allowed for API requests before considering them unsuccessful.
Default value: 3

Boolean output <N> (BOOLEAN_OUT_<N>)

No

String

Descriptor for the Nth boolean output in the structured output generated by the enricher, providing a description for the extracted boolean data (from 1 to 5).

Date output <N> (DATE_OUT_<N>)

No

String

Descriptor for the Nth date output in the structured output generated by the enricher, providing a description for the extracted date data (from 1 to 5).

Number output <N> (NUMBER_OUT_<N>)

No

String

Descriptor for the Nth number output in the structured output generated by the enricher, providing a description for the extracted number data (from 1 to 5).

String output <N> (STRING_OUT_<N>)

No

String

Descriptor for the Nth string output in the structured output generated by the enricher, providing a description for the extracted string data (from 1 to 5).
The enricher can return up to 20 outputs (five of each type).
The output descriptors are specifically designed to match the corresponding attribute types, whether they are dates, strings, numbers, or boolean values. For example, the Date output 1 descriptor exclusively matches date attributes. This matching process is automatically handled by the plugin.

Language models

Language models are AI systems trained on vast amounts of text data to understand and generate human-like language, enabling tasks like text completion, translation, summarization, and sentiment analysis.

The Vertex AI API offers a range of models with distinct capabilities. For more information on Gemini models and a list of stable model versions, see the official Vertex AI documentation.

Tokens

Tokens are units of text that language models use to process and generate language. They can range from individual characters to entire words, depending on the language and the specific model being used.

For more information about tokens, see the official Vertex AI documentation.

Plugin inputs

The following table lists the plugin inputs.

Input name Mandatory Type Description

User Prompt

No

String

Instructions specifying the information to be extracted and the method for structuring the outputs accordingly.

Source Text for Extraction

No

String

Unstructured text from which structured data is extracted.

System Prompt

No

String

Initial instruction designed to guide the model towards specific topics, styles, tones, or formats of generated text.
If you choose not to set a user prompt, you must enter a source text for extraction, and vice-versa. Defining either a user prompt or a source text for extraction is mandatory.

When opting for the Source Text for Extraction method, the enricher injects the provided text along with the configured descriptors (i.e., String output 1, Number output 1, etc.) into a standard user prompt. The pieces of information specified by the descriptors are then extracted from the text content.

When opting for the User Prompt method, model designers construct a user prompt containing unstructured values and output keys (i.e., STRING_OUT_1, NUMBER_OUT_1, etc.). These keys are then mapped to the relevant attributes in the plugin output properties.

When formulating a user prompt, make sure to instruct the enricher to extract data in a structured JSON format.

For a detailed demonstration of these methods, see Examples and use cases.

Plugin outputs

The following table lists the plugin outputs.

Output name Type Description

Boolean output <N> (BOOLEAN_OUT_<N>)

String

Extracted boolean corresponding to the Nth boolean output descriptor, numbered from 1 to 5, and applied to a designated attribute

Date output <N> (DATE_OUT_<N>)

String

Extracted date corresponding to the Nth date output descriptor, numbered from 1 to 5, and applied to a designated attribute.

Number output <N> (NUMBER_OUT_<N>)

String

Extracted number corresponding to the Nth number output descriptor, numbered from 1 to 5, and applied to a designated attribute.

String output <N> (STRING_OUT_<N>)

String

Extracted string corresponding to the Nth string output descriptor, numbered from 1 to 5, and applied to a designated attribute.

Examples and use cases

AI-powered product data extraction: streamlining record creation

Imagine a scenario where a user wants to expedite product record creation by automatically extracting a product’s name, price, and country of origin from a detailed description. In practice, the user wants the Product Name, Price, and Country of Origin fields to be automatically populated based on the Description field’s content.

For instance, consider a new product record with the following description:
"The Aerodynamic Helmet by Velocity Bikes is expertly crafted in France for speed, style, and safety. Its sleek profile reduces drag while prioritizing rider protection. Priced at $129.99, it’s the ultimate choice for safety-conscious cyclists."

Two methods can be employed to achieve the desired result.

  • Using the Source Text for Extraction method, a model designer can configure the enricher as follows:

    • In the plugin properties:

      • String output 1 (STRING_OUT_1): Name of the product

      • String output 2 (STRING_OUT_2): Country of origin of the product

      • Number output 1 (NUMBER_OUT_1): Price of the product

    • In the plugin input properties:

      • Source Text for Extraction: Description

    • In the plugin output properties:

      • ProductName: String output 1 (STRING_OUT_1)

      • Origin: String output 2 (STRING_OUT_2)

      • Price: Number output 1 (NUMBER_OUT_1)

  • Using the User Prompt method, the model designer can configure the enricher as follows:

    • In the plugin input properties:

      • User Prompt: 'From ' || Description || ' extract the following information in a structured JSON format: STRING_OUT_1: the product name, STRING_OUT_2: the country of origin, NUMBER_OUT_1: the product price.'

    • In the plugin output properties:

      • ProductName: String output 1 (STRING_OUT_1)

      • Origin: String output 2 (STRING_OUT_2)

      • Price: Number output 1 (NUMBER_OUT_1)

Regardless of the method selected, the enricher’s response populates the Product Name, Country of Origin, and Price fields in the new record with the following information:

  • Product Name: Aerodynamic Helmet

  • Country of Origin: France

  • Price: 129.99