Filter data in Azure Open AI Completion Extensions by content type

Miguel Lopes Isidoro 41

Hi,

I am using the Azure Open AI Completition Extensions in a project I am in.

Basically, I have a Azure Blob Storage with a container with the following types of information (files both in PDF and Markdown formats):

Generic articles
Events
Documents
People (Name, Email, Mobile Phone, Skills, Past Projects)

I then used the "Add Your Own Data" feature to create a data source, index and indexer that indexes the content that is target in the Chat Completions API calls.

I am having satisfactory results to every prompt except for people related questions. If I ask questions like the ones below the results return a lot of information (mostly irrelevant) but not People results.

Who is Miguel Isidoro?
Who has expertise in TypeScript?

What I would like to do is the following:

Have a way in the Blob Storage to distinguish between the above content types
Have an index with a content type field
Have a way for the indexing process to populate this new field according to the content type
Have a way in the Completion Extensions API to filter by a content type

I would like to continue using the Completion Extensions API.

Is this possible? If yes, how?

If not, what is the alternative to be able to filter by a specific content type?

Thanks,

Miguel

1 answer

Janarthanan S 700

Hi @Miguel Lopes Isidoro

Yes it is possible to Azure OpenAI on your data lets you restrict the documents that can be used in responses for different users with Azure Cognitive Search security filters. When you enable document level access, the search results returned from Azure Cognitive Search and used to generate a response will be trimmed based on user Microsoft Entra group membership. You can only enable document-level access on existing Azure Cognitive search indexes.

You can find it in the detailed documentation as mentioned below Document level access control.

https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/use-your-data

The content filtering system integrated into Azure OpenAI Service runs alongside the core models and uses an ensemble of multi-class classification models to detect four categories of harmful content (violence, hate, sexual, and self-harm) at four severity levels respectively (safe, low, medium, and high).

Please find content filtering system in the Microsoft documentation.

https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/content-filters

I hope the solution is useful to you and then accept the answer.

Regards,

Janarthanan S

Miguel Lopes Isidoro 41 Reputation points

2023-10-28T10:26:23.1166667+00:00
Hello,

Thanks for the reply.

I don't want to security trim my search results, although it is useful in other scenarios.

My scenario is the following which is related to filter data but based on metadata in the index and not filter based on AAD identities.

I am using the Azure Open AI Completition Extensions in a project I am in.

Basically, I have a Azure Blob Storage with a container with the following types of information (files both in PDF and Markdown formats):

Generic articles

Events

Documents

People (Name, Email, Mobile Phone, Skills, Past Projects)

I am having satisfactory results to every prompt except for people related questions. If I ask questions like the ones below the results return a lot of information (mostly irrelevant) but not People results.

Who is Miguel Isidoro?

Who has expertise in TypeScript?

What I would like to do is the following:

Have a way in the Blob Storage to distinguish between the above content types

Have an index with a content type field

Have a way for the indexing process to populate this new field according to the content type

Have a way in the Completion Extensions API to filter by a content type

I would like to continue using the Completion Extensions API.

Is this possible? If yes, how?

If not, what is the alternative to be able to filter by a specific content type?

Thanks,

Miguel

Janarthanan S 700

Hi @Miguel Lopes Isidoro

Please find code for metadata in the index completion playground as mentioned below.

import openai
import os
from dotenv import load_dotenv
import chromadb
from typing import List
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field
import json
load_dotenv()
openai.api_key = os.getenv("OPENAI_API_KEY")

# initialize chroma with sample data 
chroma_client = chromadb.Client()
collection = chroma_client.create_collection("my_collection")
sample_data = [
    {
        "document": "OpenAI’s ‘Dota 2’ neural nets are defeating human opponents",
        "metadata": {
            "year": 2018,
            "source": "Tech Crunch"
        },
        "id": "1"
    },
    {
        "document": "OpenAI Plans to Up the Ante in Tech’s A.I. Race",
        "metadata": {
            "year":  2013,
            "source": "The New York Times"
        },
        "id": "2"
    }
]
collection.add(
        ids=[data["id"] for data in sample_data],
        documents=[data["document"] for data in sample_data],
        metadatas=[data["metadata"] for data in sample_data]
    )

# create filter
class Filter(BaseModel):
    year : int = Field(..., description="Year of publication")
    source: str = Field(..., description="News article source")
FILTER_SCHEMA = Filter.schema()
FILTER_API = {
    "name": "filter",
    "parameters": FILTER_SCHEMA
}

# query openai for fuction using filter
query = "Find an article about openai from 2018 written by Tech Crunch"
completion = openai.ChatCompletion.create(
        model="gpt-3.5-turbo-0613",
        temperature=0,
        functions=[FILTER_API],
        messages=[
            {
                "role": "system",
                "content": "You must use the `metadata` function .",
            },
            {
                "role": "user",
                "content": query
            },
        ],
    )
filter = json.loads(completion.choices[0].message.function_call.arguments)
print('Query:', query)
print('Metadata filter:',  filter)
dbquery = collection.query(
    n_results=1,
    query_texts=query,
    where=filter 
)
print('Query Result:', dbquery["documents"][0][0])

I hope the solution is useful to you and then accept the answer.

Regards,

Janarthanan S

Miguel Lopes Isidoro 41 Reputation points

2023-10-28T16:44:26.6633333+00:00

Hello,

Is it possible to send the code in C#?

Thanks

Janarthanan S 700

using System;
using System.Collections.Generic;
using OpenAI;
using System.IO;
using dotenv.net;

public class Program
{
    public static void Main()
    {
        DotEnv.Config();
        OpenAI.Configuration.ApiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY");

        // initialize chroma with sample data 
        var chromaClient = new Chromadb.Client();
        var collection = chromaClient.CreateCollection("my_collection");
        var sampleData = new List<Dictionary<string, object>>
        {
            new Dictionary<string, object>
            {
                { "document", "OpenAI’s ‘Dota 2’ neural nets are defeating human opponents" },
                { "metadata", new Dictionary<string, object>
                    {
                        { "year", 2018 },
                        { "source", "Tech Crunch" }
                    }
                },
                { "id", "1" }
            },
            new Dictionary<string, object>
            {
                { "document", "OpenAI Plans to Up the Ante in Tech’s A.I. Race" },
                { "metadata", new Dictionary<string, object>
                    {
                        { "year", 2013 },
                        { "source", "The New York Times" }
                    }
                },
                { "id", "2" }
            }
        };
        collection.Add(
            ids: sampleData.ConvertAll(data => data["id"].ToString()),
            documents: sampleData.ConvertAll(data => data["document"].ToString()),
            metadatas: sampleData.ConvertAll(data => (Dictionary<string, object>)data["metadata"])
        );

        // create filter
        public class Filter
        {
            public int year { get; set; }
            public string source { get; set; }
        }
        var filter = new Filter
        {
            year = 2018,
            source = "Tech Crunch"
        };

        // query openai for function using filter
        var query = "Find an article about openai from 2018 written by Tech Crunch";
        var completion = OpenAI.ChatCompletion.Create(
            model: "gpt-3.5-turbo-0613",
            temperature: 0,
            functions: new List<Dictionary<string, object>>
            {
                new Dictionary<string, object>
                {
                    { "name", "filter" },
                    { "parameters", filter }
                }
            },
            messages: new List<Dictionary<string, string>>
            {
                new Dictionary<string, string>
                {
                    { "role", "system" },
                    { "content", "You must use the `metadata` function ." }
                },
                new Dictionary<string, string>
                {
                    { "role", "user" },
                    { "content", query }
                }
            }
        );
        var filterResult = (Dictionary<string, object>)completion.Choices[0].Message.FunctionCall.Arguments;
        Console.WriteLine("Query: " + query);
        Console.WriteLine("Metadata filter: " + filterResult);
        var dbquery = collection.Query(
            nResults: 1,
            queryTexts: new List<string> { query },
            where: filterResult
        );
        Console.WriteLine("Query Result: " + dbquery["documents"][0][0]);
    }
}

Hi @Miguel Lopes Isidoro

Please find C# Code mentioned above.

I hope the solution is useful to you and then accept the answer.

Regards,

Janarthanan S

Miguel Lopes Isidoro 41 Reputation points

2023-10-29T09:51:59.65+00:00
Hi,

This code doesn't address my scenario.

My scenario is the following which is related to filter data but based on metadata in the index and not filter based on AAD identities and using the Completions Extensions API.

Basically, I have a Azure Blob Storage with a container with the following types of information (files both in PDF and Markdown formats):

Generic articles

Events

Documents

People (Name, Email, Mobile Phone, Skills, Past Projects)

I am having satisfactory results to every prompt except for people related questions. If I ask questions like the ones below the results return a lot of information (mostly irrelevant) but not People results.

Who is Miguel Isidoro?

Who has expertise in TypeScript?

What I would like to do is the following:

Have a way in the Blob Storage to distinguish between the above content types

Have an index with a content type field

Have a way for the indexing process to populate this new field according to the content type

Have a way in the Completion Extensions API to filter by a content type

I would like to continue using the Completion Extensions API.

Is this possible? If yes, how?

If not, what is the alternative to be able to filter by a specific content type?

Thanks,

Miguel

Share via

Filter data in Azure Open AI Completion Extensions by content type

1 answer

Your answer