Add an optional seed parameter (#1814)

* Add seed parameter

* Assert seed is None in tests

* Add tests

* Add tests snapshots

Author: pamelafox

app/backend/approaches/chatreadretrieveread.py

@@ -88,6 +88,7 @@ async def run_until_final_call(
         auth_claims: dict[str, Any],
         should_stream: bool = False,
     ) -> tuple[dict[str, Any], Coroutine[Any, Any, Union[ChatCompletion, AsyncStream[ChatCompletionChunk]]]]:
+        seed = overrides.get("seed", None)
         use_text_search = overrides.get("retrieval_mode") in ["text", "hybrid", None]
         use_vector_search = overrides.get("retrieval_mode") in ["vectors", "hybrid", None]
         use_semantic_ranker = True if overrides.get("semantic_ranker") else False
@@ -142,6 +143,7 @@ async def run_until_final_call(
             max_tokens=query_response_token_limit,  # Setting too low risks malformed JSON, setting too high may affect performance
             n=1,
             tools=tools,
+            seed=seed,
         )
 
         query_text = self.get_search_query(chat_completion, original_user_query)
@@ -237,5 +239,6 @@ async def run_until_final_call(
             max_tokens=response_token_limit,
             n=1,
             stream=should_stream,
+            seed=seed,
         )
         return (extra_info, chat_coroutine)

app/backend/approaches/chatreadretrievereadvision.py

@@ -89,6 +89,7 @@ async def run_until_final_call(
         auth_claims: dict[str, Any],
         should_stream: bool = False,
     ) -> tuple[dict[str, Any], Coroutine[Any, Any, Union[ChatCompletion, AsyncStream[ChatCompletionChunk]]]]:
+        seed = overrides.get("seed", None)
         use_text_search = overrides.get("retrieval_mode") in ["text", "hybrid", None]
         use_vector_search = overrides.get("retrieval_mode") in ["vectors", "hybrid", None]
         use_semantic_ranker = True if overrides.get("semantic_ranker") else False
@@ -128,6 +129,7 @@ async def run_until_final_call(
             temperature=0.0,  # Minimize creativity for search query generation
             max_tokens=query_response_token_limit,
             n=1,
+            seed=seed,
         )
 
         query_text = self.get_search_query(chat_completion, original_user_query)
@@ -241,5 +243,6 @@ async def run_until_final_call(
             max_tokens=response_token_limit,
             n=1,
             stream=should_stream,
+            seed=seed,
         )
         return (extra_info, chat_coroutine)

app/backend/approaches/retrievethenread.py

@@ -79,6 +79,7 @@ async def run(
         if not isinstance(q, str):
             raise ValueError("The most recent message content must be a string.")
         overrides = context.get("overrides", {})
+        seed = overrides.get("seed", None)
         auth_claims = context.get("auth_claims", {})
         use_text_search = overrides.get("retrieval_mode") in ["text", "hybrid", None]
         use_vector_search = overrides.get("retrieval_mode") in ["vectors", "hybrid", None]
@@ -131,6 +132,7 @@ async def run(
                 temperature=overrides.get("temperature", 0.3),
                 max_tokens=response_token_limit,
                 n=1,
+                seed=seed,
             )
         ).model_dump()
 

app/backend/approaches/retrievethenreadvision.py

@@ -80,6 +80,7 @@ async def run(
             raise ValueError("The most recent message content must be a string.")
 
         overrides = context.get("overrides", {})
+        seed = overrides.get("seed", None)
         auth_claims = context.get("auth_claims", {})
         use_text_search = overrides.get("retrieval_mode") in ["text", "hybrid", None]
         use_vector_search = overrides.get("retrieval_mode") in ["vectors", "hybrid", None]
@@ -148,6 +149,7 @@ async def run(
                 temperature=overrides.get("temperature", 0.3),
                 max_tokens=response_token_limit,
                 n=1,
+                seed=seed,
             )
         ).model_dump()
 

app/frontend/src/api/models.ts

@@ -21,6 +21,7 @@ export type ChatAppRequestOverrides = {
     semantic_ranker?: boolean;
     semantic_captions?: boolean;
     exclude_category?: string;
+    seed?: number;
     top?: number;
     temperature?: number;
     minimum_search_score?: number;

app/frontend/src/i18n/tooltips.ts

@@ -5,6 +5,7 @@ export const toolTipText = {
         "Overrides the prompt used to generate the answer based on the question and search results. To append to existing prompt instead of replace whole prompt, start your prompt with '>>>'.",
     temperature:
         "Sets the temperature of the request to the LLM that generates the answer. Higher temperatures result in more creative responses, but they may be less grounded.",
+    seed: "Sets a seed to improve the reproducibility of the model's responses. The seed can be any integer.",
     searchScore:
         "Sets a minimum score for search results coming back from Azure AI search. The score range depends on whether you're using hybrid (default), vectors only, or text only.",
     rerankerScore:

app/frontend/src/pages/ask/Ask.tsx

@@ -27,6 +27,7 @@ export function Component(): JSX.Element {
     const [promptTemplatePrefix, setPromptTemplatePrefix] = useState<string>("");
     const [promptTemplateSuffix, setPromptTemplateSuffix] = useState<string>("");
     const [temperature, setTemperature] = useState<number>(0.3);
+    const [seed, setSeed] = useState<number | null>(null);
     const [minimumRerankerScore, setMinimumRerankerScore] = useState<number>(0);
     const [minimumSearchScore, setMinimumSearchScore] = useState<number>(0);
     const [retrievalMode, setRetrievalMode] = useState<RetrievalMode>(RetrievalMode.Hybrid);
@@ -124,7 +125,8 @@ export function Component(): JSX.Element {
                         use_groups_security_filter: useGroupsSecurityFilter,
                         vector_fields: vectorFieldList,
                         use_gpt4v: useGPT4V,
-                        gpt4v_input: gpt4vInput
+                        gpt4v_input: gpt4vInput,
+                        ...(seed !== null ? { seed: seed } : {})
                     }
                 },
                 // AI Chat Protocol: Client must pass on any session state received from the server
@@ -148,6 +150,10 @@ export function Component(): JSX.Element {
         setTemperature(parseFloat(newValue || "0"));
     };
 
+    const onSeedChange = (_ev?: React.SyntheticEvent<HTMLElement, Event>, newValue?: string) => {
+        setSeed(parseInt(newValue || ""));
+    };
+
     const onMinimumSearchScoreChange = (_ev?: React.SyntheticEvent<HTMLElement, Event>, newValue?: string) => {
         setMinimumSearchScore(parseFloat(newValue || "0"));
     };
@@ -206,6 +212,8 @@ export function Component(): JSX.Element {
     const promptTemplateFieldId = useId("promptTemplateField");
     const temperatureId = useId("temperature");
     const temperatureFieldId = useId("temperatureField");
+    const seedId = useId("seed");
+    const seedFieldId = useId("seedField");
     const searchScoreId = useId("searchScore");
     const searchScoreFieldId = useId("searchScoreField");
     const rerankerScoreId = useId("rerankerScore");
@@ -314,6 +322,19 @@ export function Component(): JSX.Element {
                     )}
                 />
 
+                <TextField
+                    id={seedFieldId}
+                    className={styles.chatSettingsSeparator}
+                    label="Seed"
+                    type="text"
+                    defaultValue={seed?.toString() || ""}
+                    onChange={onSeedChange}
+                    aria-labelledby={seedId}
+                    onRenderLabel={(props: ITextFieldProps | undefined) => (
+                        <HelpCallout labelId={seedId} fieldId={seedFieldId} helpText={toolTipText.seed} label={props?.label} />
+                    )}
+                />
+
                 <TextField
                     id={searchScoreFieldId}
                     className={styles.chatSettingsSeparator}

app/frontend/src/pages/chat/Chat.tsx

@@ -39,6 +39,7 @@ const Chat = () => {
     const [isConfigPanelOpen, setIsConfigPanelOpen] = useState(false);
     const [promptTemplate, setPromptTemplate] = useState<string>("");
     const [temperature, setTemperature] = useState<number>(0.3);
+    const [seed, setSeed] = useState<number | null>(null);
     const [minimumRerankerScore, setMinimumRerankerScore] = useState<number>(0);
     const [minimumSearchScore, setMinimumSearchScore] = useState<number>(0);
     const [retrieveCount, setRetrieveCount] = useState<number>(3);
@@ -173,7 +174,8 @@ const Chat = () => {
                         use_groups_security_filter: useGroupsSecurityFilter,
                         vector_fields: vectorFieldList,
                         use_gpt4v: useGPT4V,
-                        gpt4v_input: gpt4vInput
+                        gpt4v_input: gpt4vInput,
+                        ...(seed !== null ? { seed: seed } : {})
                     }
                 },
                 // AI Chat Protocol: Client must pass on any session state received from the server
@@ -239,6 +241,10 @@ const Chat = () => {
         setTemperature(parseFloat(newValue || "0"));
     };
 
+    const onSeedChange = (_ev?: React.SyntheticEvent<HTMLElement, Event>, newValue?: string) => {
+        setSeed(parseInt(newValue || ""));
+    };
+
     const onMinimumSearchScoreChange = (_ev?: React.SyntheticEvent<HTMLElement, Event>, newValue?: string) => {
         setMinimumSearchScore(parseFloat(newValue || "0"));
     };
@@ -309,6 +315,8 @@ const Chat = () => {
     const promptTemplateFieldId = useId("promptTemplateField");
     const temperatureId = useId("temperature");
     const temperatureFieldId = useId("temperatureField");
+    const seedId = useId("seed");
+    const seedFieldId = useId("seedField");
     const searchScoreId = useId("searchScore");
     const searchScoreFieldId = useId("searchScoreField");
     const rerankerScoreId = useId("rerankerScore");
@@ -478,6 +486,19 @@ const Chat = () => {
                         )}
                     />
 
+                    <TextField
+                        id={seedFieldId}
+                        className={styles.chatSettingsSeparator}
+                        label="Seed"
+                        type="text"
+                        defaultValue={seed?.toString() || ""}
+                        onChange={onSeedChange}
+                        aria-labelledby={seedId}
+                        onRenderLabel={(props: ITextFieldProps | undefined) => (
+                            <HelpCallout labelId={seedId} fieldId={seedFieldId} helpText={toolTipText.seed} label={props?.label} />
+                        )}
+                    />
+
                     <TextField
                         id={searchScoreFieldId}
                         className={styles.chatSettingsSeparator}

tests/conftest.py

@@ -180,6 +180,9 @@ async def __anext__(self):
                 raise StopAsyncIteration
 
     async def mock_acreate(*args, **kwargs):
+        # The only two possible values for seed:
+        assert kwargs.get("seed") is None or kwargs.get("seed") == 42
+
         messages = kwargs["messages"]
         last_question = messages[-1]["content"]
         if last_question == "Generate search query for: What is the capital of France?":

tests/snapshots/test_app/test_chat_seed/client0/result.json

@@ -0,0 +1,83 @@
+{
+    "context": {
+        "data_points": {
+            "text": [
+                "Benefit_Options-2.pdf: There is a whistleblower policy."
+            ]
+        },
+        "thoughts": [
+            {
+                "description": [
+                    "{'role': 'system', 'content': \"Below is a history of the conversation so far, and a new question asked by the user that needs to be answered by searching in a knowledge base.\\n    You have access to Azure AI Search index with 100's of documents.\\n    Generate a search query based on the conversation and the new question.\\n    Do not include cited source filenames and document names e.g info.txt or doc.pdf in the search query terms.\\n    Do not include any text inside [] or <<>> in the search query terms.\\n    Do not include any special characters like '+'.\\n    If the question is not in English, translate the question to English before generating the search query.\\n    If you cannot generate a search query, return just the number 0.\\n    \"}",
+                    "{'role': 'user', 'content': 'How did crypto do last year?'}",
+                    "{'role': 'assistant', 'content': 'Summarize Cryptocurrency Market Dynamics from last year'}",
+                    "{'role': 'user', 'content': 'What are my health plans?'}",
+                    "{'role': 'assistant', 'content': 'Show available health plans'}",
+                    "{'role': 'user', 'content': 'Generate search query for: What is the capital of France?'}"
+                ],
+                "props": {
+                    "model": "gpt-35-turbo"
+                },
+                "title": "Prompt to generate search query"
+            },
+            {
+                "description": "capital of France",
+                "props": {
+                    "filter": null,
+                    "top": 3,
+                    "use_semantic_captions": false,
+                    "use_semantic_ranker": false,
+                    "use_text_search": true,
+                    "use_vector_search": true
+                },
+                "title": "Search using generated search query"
+            },
+            {
+                "description": [
+                    {
+                        "captions": [
+                            {
+                                "additional_properties": {},
+                                "highlights": [],
+                                "text": "Caption: A whistleblower policy."
+                            }
+                        ],
+                        "category": null,
+                        "content": "There is a whistleblower policy.",
+                        "embedding": null,
+                        "groups": null,
+                        "id": "file-Benefit_Options_pdf-42656E656669745F4F7074696F6E732E706466-page-2",
+                        "imageEmbedding": null,
+                        "oids": null,
+                        "reranker_score": 3.4577205181121826,
+                        "score": 0.03279569745063782,
+                        "sourcefile": "Benefit_Options.pdf",
+                        "sourcepage": "Benefit_Options-2.pdf"
+                    }
+                ],
+                "props": null,
+                "title": "Search results"
+            },
+            {
+                "description": [
+                    "{'role': 'system', 'content': \"Assistant helps the company employees with their healthcare plan questions, and questions about the employee handbook. Be brief in your answers.\\n        Answer ONLY with the facts listed in the list of sources below. If there isn't enough information below, say you don't know. Do not generate answers that don't use the sources below. If asking a clarifying question to the user would help, ask the question.\\n        For tabular information return it as an html table. Do not return markdown format. If the question is not in English, answer in the language used in the question.\\n        Each source has a name followed by colon and the actual information, always include the source name for each fact you use in the response. Use square brackets to reference the source, for example [info1.txt]. Don't combine sources, list each source separately, for example [info1.txt][info2.pdf].\\n        \\n        \\n        \"}",
+                    "{'role': 'user', 'content': 'What is the capital of France?\\n\\nSources:\\nBenefit_Options-2.pdf: There is a whistleblower policy.'}"
+                ],
+                "props": {
+                    "model": "gpt-35-turbo"
+                },
+                "title": "Prompt to generate answer"
+            }
+        ]
+    },
+    "finish_reason": "stop",
+    "index": 0,
+    "logprobs": null,
+    "message": {
+        "content": "The capital of France is Paris. [Benefit_Options-2.pdf].",
+        "function_call": null,
+        "role": "assistant",
+        "tool_calls": null
+    },
+    "session_state": null
+}