Add more request options

jelledruyts · jelledruyts · commit bdbe00017e4d · 2023-12-19T16:17:08.000+01:00
diff --git a/src/Azure.AISearch.WebApp/Constants.cs b/src/Azure.AISearch.WebApp/Constants.cs
@@ -58,6 +58,8 @@ public static class Defaults
         public const double FrequencyPenalty = 0.0;
         public const double PresencePenalty = 0.0;
         public const string StopSequences = "";
+        public const int Strictness = 3;
+        public const int DocumentCount = 5;
         public const int HnswParametersM = 4;
         public const int HnswParametersEfConstruction = 400;
         public const int HnswParametersEfSearch = 500;
diff --git a/src/Azure.AISearch.WebApp/Models/SearchRequest.cs b/src/Azure.AISearch.WebApp/Models/SearchRequest.cs
@@ -21,6 +21,8 @@ public class SearchRequest
     public double? FrequencyPenalty { get; set; } = Constants.Defaults.FrequencyPenalty;
     public double? PresencePenalty { get; set; } = Constants.Defaults.PresencePenalty;
     public string? StopSequences { get; set; } = Constants.Defaults.StopSequences;
+    public int? Strictness { get; set; } = Constants.Defaults.Strictness;
+    public int? DocumentCount { get; set; } = Constants.Defaults.DocumentCount;
 
     public bool IsVectorSearch => QueryType == Models.QueryType.Vector || QueryType == Models.QueryType.HybridStandard || QueryType == Models.QueryType.HybridSemantic;
     public bool IsSemanticSearch => QueryType == Models.QueryType.TextSemantic || QueryType == Models.QueryType.HybridSemantic;
diff --git a/src/Azure.AISearch.WebApp/Pages/Index.cshtml b/src/Azure.AISearch.WebApp/Pages/Index.cshtml
@@ -274,8 +274,8 @@
                         <label class="form-check-label" for="searchRequest-DataSource-AzureCognitiveSearch">Azure AI Search</label>
                         <span class="info-tip" data-bs-toggle="popover" data-bs-content="Use source data from an Azure AI Search index. For best results, you should use the <code>Chunks</code> index, as this contains smaller and typically more contextually relevant pieces of information. This makes the prompt that is sent to the AI model better suited to generate a meaningful response from."><i class="bi bi-info-circle"></i></span>
                     </div>
-                    <div class="form-check form-check-inline form-switch">
-                        <input class="form-check-input" type="checkbox" role="switch" name="@nameof(SearchRequest.LimitToDataSource)" id="searchRequest-LimitToDataSource" value="true" v-model="searchRequest.limitToDataSource" v-bind:disabled="searchRequest.dataSource == '@DataSourceType.None'">
+                    <div class="form-check form-check-inline form-switch" v-show="searchRequest.dataSource != '@DataSourceType.None'">
+                        <input class="form-check-input" type="checkbox" role="switch" name="@nameof(SearchRequest.LimitToDataSource)" id="searchRequest-LimitToDataSource" value="true" v-model="searchRequest.limitToDataSource">
                         <input type="hidden" name="@nameof(SearchRequest.LimitToDataSource)" value="false" />
                         <label class="form-check-label" for="searchRequest-LimitToDataSource">Limit to your data</label>
                         <span class="info-tip" data-bs-toggle="popover" data-bs-content="Limit responses to your data content only, not including the data that was used for training the model."><i class="bi bi-info-circle"></i></span>
@@ -312,18 +312,18 @@
                         <label class="form-check-label" for="searchRequest-QueryType-TextSemantic">Semantic keyword</label>
                         <span class="info-tip" data-bs-toggle="popover" data-bs-content="Use <a href='https://learn.microsoft.com/azure/search/semantic-search-overview' target='_blank'>semantic ranking</a>, which returns more relevant results by applying language understanding to initial search results. It can also return <a href='https://learn.microsoft.com/azure/search/semantic-how-to-query-request' target='_blank'>semantic captions</a> (parts of a document that best summarize the content) and even <a href='https://learn.microsoft.com/azure/search/semantic-answers' target='_blank'>semantic answers</a> (direct answers to queries that look like a question). In all cases, the responses aren't AI-generated but come directly from the source data."><i class="bi bi-info-circle"></i></span>
                     </div>
-                    <div class="form-check form-check-inline">
-                        <input class="form-check-input" type="radio" name="@nameof(SearchRequest.QueryType)" id="searchRequest-QueryType-Vector" value="@QueryType.Vector" v-model="searchRequest.queryType" v-bind:disabled="searchRequest.searchIndex == '@SearchIndexType.Documents'">
+                    <div class="form-check form-check-inline" v-show="searchRequest.searchIndex != '@SearchIndexType.Documents'">
+                        <input class="form-check-input" type="radio" name="@nameof(SearchRequest.QueryType)" id="searchRequest-QueryType-Vector" value="@QueryType.Vector" v-model="searchRequest.queryType">
                         <label class="form-check-label" for="searchRequest-QueryType-Vector">Vector only</label>
                         <span class="info-tip" data-bs-toggle="popover" data-bs-content="First send the search query to an <a href='https://learn.microsoft.com/azure/ai-services/openai/how-to/embeddings' target='_blank'>embedding model in Azure OpenAI</a> to generate a vector representing the query itself. Then perform a <a href='https://learn.microsoft.com/azure/search/vector-search-overview' target='_blank'>vector search</a> to retrieve the nearest neighbors in vector space from the chunked and vectorized documents in the <code>Chunks</code> index. This should return results that are semantically similar to the query, as determined by the embedding model's vector representations."><i class="bi bi-info-circle"></i></span>
                     </div>
-                    <div class="form-check form-check-inline">
-                        <input class="form-check-input" type="radio" name="@nameof(SearchRequest.QueryType)" id="searchRequest-QueryType-HybridStandard" value="@QueryType.HybridStandard" v-model="searchRequest.queryType" v-bind:disabled="searchRequest.searchIndex == '@SearchIndexType.Documents'">
+                    <div class="form-check form-check-inline" v-show="searchRequest.searchIndex != '@SearchIndexType.Documents'">
+                        <input class="form-check-input" type="radio" name="@nameof(SearchRequest.QueryType)" id="searchRequest-QueryType-HybridStandard" value="@QueryType.HybridStandard" v-model="searchRequest.queryType">
                         <label class="form-check-label" for="searchRequest-QueryType-HybridStandard">Standard hybrid (keyword + vector)</label>
                         <span class="info-tip" data-bs-toggle="popover" data-bs-content="Combine the results of a <a href='https://learn.microsoft.com/azure/search/vector-search-overview' target='_blank'>vector search</a> with the regular <a href='https://learn.microsoft.com/azure/search/search-lucene-query-architecture' target='_blank'>keyword search</a> results into a single ranked response."><i class="bi bi-info-circle"></i></span>
                     </div>
-                    <div class="form-check form-check-inline">
-                        <input class="form-check-input" type="radio" name="@nameof(SearchRequest.QueryType)" id="searchRequest-QueryType-HybridSemantic" value="@QueryType.HybridSemantic" v-model="searchRequest.queryType" v-bind:disabled="searchRequest.searchIndex == '@SearchIndexType.Documents'">
+                    <div class="form-check form-check-inline" v-show="searchRequest.searchIndex != '@SearchIndexType.Documents'">
+                        <input class="form-check-input" type="radio" name="@nameof(SearchRequest.QueryType)" id="searchRequest-QueryType-HybridSemantic" value="@QueryType.HybridSemantic" v-model="searchRequest.queryType">
                         <label class="form-check-label" for="searchRequest-QueryType-HybridSemantic">Semantic hybrid (keyword + vector)</label>
                         <span class="info-tip" data-bs-toggle="popover" data-bs-content="Combine the results of <a href='https://learn.microsoft.com/azure/search/vector-search-overview' target='_blank'>vector search</a> and <a href='https://learn.microsoft.com/azure/search/semantic-search-overview' target='_blank'>search with semantic ranking</a> into a single ranked response. Compared to standard hybrid search, this provides even more accuracy with L2 reranking using the same language models that power Bing."><i class="bi bi-info-circle"></i></span>
                     </div>
@@ -346,6 +346,30 @@
                 </div>
             </div>
 
+            <div class="mb-2" v-show="!(searchRequest.engine == '@EngineType.AzureOpenAI' && searchRequest.dataSource == '@DataSourceType.None')">
+                <label class="form-label">Result parameters</label>
+                <div class="row row-cols-lg-auto g-3 align-items-center">
+                    <div class="col-12">
+                        <div class="input-group">
+                            <div class="input-group-text">
+                                Document count
+                                <span class="info-tip" data-bs-toggle="popover" data-bs-content="The number of documents to return."><i class="bi bi-info-circle"></i></span>
+                            </div>
+                            <input type="number" class="form-control form-control-small" name="@nameof(SearchRequest.DocumentCount)" value="@Model.SearchRequest.DocumentCount" />
+                        </div>
+                    </div>
+                    <div class="col-12" v-show="searchRequest.engine == '@EngineType.AzureOpenAI' && searchRequest.dataSource == '@DataSourceType.AzureCognitiveSearch'">
+                        <div class="input-group">
+                            <div class="input-group-text">
+                                Strictness
+                                <span class="info-tip" data-bs-toggle="popover" data-bs-content="Sets the threshold to categorize documents as relevant to your queries. Raising the value means a higher threshold for relevance and filters out more less-relevant documents for responses. Setting this value too high might cause the model to fail to generate responses due to limited available documents."><i class="bi bi-info-circle"></i></span>
+                            </div>
+                            <input type="number" class="form-control form-control-small" name="@nameof(SearchRequest.Strictness)" value="@Model.SearchRequest.Strictness" />
+                        </div>
+                    </div>
+                </div>
+            </div>
+
             <div class="mb-2" v-show="!(searchRequest.engine == '@EngineType.AzureOpenAI' && searchRequest.dataSource == '@DataSourceType.None') && (searchRequest.queryType == '@QueryType.Vector' || searchRequest.queryType == '@QueryType.HybridStandard' || searchRequest.queryType == '@QueryType.HybridSemantic')">
                 <label class="form-label">Vector parameters</label>
                 <div class="row row-cols-lg-auto g-3 align-items-center">
@@ -359,8 +383,8 @@
                         </div>
                     </div>
                     <div class="col-12">
-                        <div class="form-check form-check-inline form-switch">
-                            <input class="form-check-input" type="checkbox" role="switch" name="@nameof(SearchRequest.UseIntegratedVectorization)" id="searchRequest-UseIntegratedVectorization" value="true" v-model="searchRequest.useIntegratedVectorization" v-bind:disabled="searchRequest.engine == '@EngineType.AzureOpenAI'">
+                        <div class="form-check form-check-inline form-switch" v-show="searchRequest.engine != '@EngineType.AzureOpenAI'">
+                            <input class="form-check-input" type="checkbox" role="switch" name="@nameof(SearchRequest.UseIntegratedVectorization)" id="searchRequest-UseIntegratedVectorization" value="true" v-model="searchRequest.useIntegratedVectorization">
                             <input type="hidden" name="@nameof(SearchRequest.UseIntegratedVectorization)" value="false" />
                             <label class="form-check-label" for="searchRequest-UseIntegratedVectorization">Use integrated vectorization</label>
                             <span class="info-tip" data-bs-toggle="popover" data-bs-content="Use <a href='https://learn.microsoft.com/azure/search/vector-search-integrated-vectorization' target='_blank'>integrated vectorization</a> to let Azure AI Search generate the vector embedding for the search query text, rather than the app doing that upfront and sending the vector directly to the search service."><i class="bi bi-info-circle"></i></span>
@@ -369,7 +393,7 @@
                 </div>
             </div>
 
-            <div class="mb-2" v-show="searchRequest.engine == '@EngineType.CustomOrchestration'">
+            <div class="mb-2" v-show="searchRequest.engine == '@EngineType.CustomOrchestration' || searchRequest.engine == '@EngineType.AzureOpenAI'">
                 <label class="form-label">Model parameters</label>
                 <div class="row row-cols-lg-auto g-3 align-items-center">
                     <div class="col-12">
diff --git a/src/Azure.AISearch.WebApp/Services/AzureCognitiveSearchService.cs b/src/Azure.AISearch.WebApp/Services/AzureCognitiveSearchService.cs
@@ -33,6 +33,7 @@ public async Task<SearchResponse> SearchAsync(SearchRequest request)
         var searchOptions = new SearchOptions
         {
             QueryType = request.IsSemanticSearch ? SearchQueryType.Semantic : (request.QuerySyntax == QuerySyntax.Lucene ? SearchQueryType.Full : SearchQueryType.Simple),
+            Size = request.DocumentCount ?? Constants.Defaults.DocumentCount,
             HighlightPreTag = "<mark>",
             HighlightPostTag = "</mark>"
         };
diff --git a/src/Azure.AISearch.WebApp/Services/AzureOpenAISearchService.cs b/src/Azure.AISearch.WebApp/Services/AzureOpenAISearchService.cs
@@ -28,7 +28,23 @@ public async Task<SearchResponse> SearchAsync(SearchRequest request)
         ArgumentNullException.ThrowIfNull(request.Query);
 
         var searchResponse = new SearchResponse();
-        var chatCompletionsOptions = new ChatCompletionsOptions { DeploymentName = this.settings.OpenAIGptDeployment };
+        var chatCompletionsOptions = new ChatCompletionsOptions
+        {
+            DeploymentName = this.settings.OpenAIGptDeployment,
+            MaxTokens = request.MaxTokens ?? Constants.Defaults.MaxTokens,
+            Temperature = (float)(request.Temperature ?? Constants.Defaults.Temperature),
+            NucleusSamplingFactor = (float)(request.TopP ?? Constants.Defaults.TopP),
+            FrequencyPenalty = (float)(request.FrequencyPenalty ?? Constants.Defaults.FrequencyPenalty),
+            PresencePenalty = (float)(request.PresencePenalty ?? Constants.Defaults.PresencePenalty),
+        };
+        var stopSequences = (request.StopSequences ?? Constants.Defaults.StopSequences).Split(',', StringSplitOptions.RemoveEmptyEntries | StringSplitOptions.TrimEntries);
+        if (stopSequences.Any())
+        {
+            foreach (var stopSequence in stopSequences)
+            {
+                chatCompletionsOptions.StopSequences.Add(stopSequence);
+            }
+        }
         chatCompletionsOptions.Messages.Add(new ChatRequestSystemMessage(request.SystemRoleInformation));
 
         if (request.History != null && request.History.Any())
@@ -66,27 +82,30 @@ public async Task<SearchResponse> SearchAsync(SearchRequest request)
             throw new InvalidOperationException("Azure OpenAI didn't return a meaningful response.");
         }
 
-        // Process citations within the answer, which take the form "[doc1][doc2]..." and refer to the (1-based) index of
-        // the citations in the tool message.
-        foreach (var extensionMessage in answerMessage.AzureExtensionsContext.Messages.Where(m => m.Role == ChatRole.Tool))
+        if (answerMessage.AzureExtensionsContext != null)
         {
-            Console.WriteLine(extensionMessage.Content);
-            var content = JsonSerializer.Deserialize<ChatResponseMessageContent>(extensionMessage.Content!);
-            if (content?.Citations != null && content.Citations.Any())
+            // Process citations within the answer, which take the form "[doc1][doc2]..." and refer to the (1-based) index of
+            // the citations in the tool message.
+            foreach (var extensionMessage in answerMessage.AzureExtensionsContext.Messages.Where(m => m.Role == ChatRole.Tool))
             {
-                var citationIndex = 0;
-                foreach (var citation in content.Citations)
+                Console.WriteLine(extensionMessage.Content);
+                var content = JsonSerializer.Deserialize<ChatResponseMessageContent>(extensionMessage.Content!);
+                if (content?.Citations != null && content.Citations.Any())
                 {
-                    answerText = answerText.Replace($"[doc{++citationIndex}]", $"<cite>{citation.Title}</cite>", StringComparison.OrdinalIgnoreCase);
-                    searchResponse.SearchResults.Add(new SearchResult
+                    var citationIndex = 0;
+                    foreach (var citation in content.Citations)
                     {
-                        DocumentId = citation.Id,
-                        DocumentTitle = citation.Title,
-                        Captions = string.IsNullOrWhiteSpace(citation.Content) ? Array.Empty<string>() : new[] { citation.Content }
-                    });
+                        answerText = answerText.Replace($"[doc{++citationIndex}]", $"<cite>{citation.Title}</cite>", StringComparison.OrdinalIgnoreCase);
+                        searchResponse.SearchResults.Add(new SearchResult
+                        {
+                            DocumentId = citation.Id,
+                            DocumentTitle = citation.Title,
+                            Captions = string.IsNullOrWhiteSpace(citation.Content) ? Array.Empty<string>() : new[] { citation.Content }
+                        });
+                    }
+                    // Stop looping through the tool messages once we find the first one holding the citations.
+                    break;
                 }
-                // Stop looping through the tool messages once we find the first one holding the citations.
-                break;
             }
         }
 
@@ -117,6 +136,9 @@ private AzureCognitiveSearchChatExtensionConfiguration GetAzureCognitiveSearchDa
             },
             ShouldRestrictResultScope = request.LimitToDataSource, // Limit responses to data from the data source only
             QueryType = GetQueryType(request),
+            RoleInformation = request.SystemRoleInformation ?? Constants.Defaults.SystemRoleInformation,
+            Strictness = request.Strictness ?? Constants.Defaults.Strictness,
+            DocumentCount = request.DocumentCount ?? Constants.Defaults.DocumentCount,
             SemanticConfiguration = request.IsSemanticSearch ? Constants.ConfigurationNames.SemanticConfigurationNameDefault : null,
             EmbeddingEndpoint = request.IsVectorSearch ? new Uri(new Uri(this.settings.OpenAIEndpoint), $"openai/deployments/{this.settings.OpenAIEmbeddingDeployment}/embeddings?api-version={this.settings.OpenAIApiVersion}") : null,
             EmbeddingKey = request.IsVectorSearch ? this.settings.OpenAIApiKey : null

Original file line number	Diff line number	Diff line change
`@@ -33,6 +33,7 @@ public async Task<SearchResponse> SearchAsync(SearchRequest request)`
`33`	`33`	`var searchOptions = new SearchOptions`
`34`	`34`	`{`
`35`	`35`	`QueryType = request.IsSemanticSearch ? SearchQueryType.Semantic : (request.QuerySyntax == QuerySyntax.Lucene ? SearchQueryType.Full : SearchQueryType.Simple),`
	`36`	`+ Size = request.DocumentCount ?? Constants.Defaults.DocumentCount,`
`36`	`37`	`HighlightPreTag = "<mark>",`
`37`	`38`	`HighlightPostTag = "</mark>"`
`38`	`39`	`};`