Alpha parameter in hybrid search is not working for azure aisearch llama-index integration

[madtur]

I am using existing index from Azure AIsearch, and all necessary fields are searchable. When I run "sparse" or "default" mode search, it gives relevant results. However, when I specify "hybrid" search, it gives me exactly the same results as in case of "default" mode search. When I tried to change alpha from 0 to 1, even invalid values like string "test" - it still gives me the same output. My thought is that in azure aisearch integration code, when search mode is hybrid, it doesn't use alpha value, that's why it is ignored. Below is part of the code from llama-index azure aisearch integration, where hybrid search is handled:

class AzureQueryResultSearchHybrid(
    AzureQueryResultSearchDefault, AzureQueryResultSearchSparse
):
    def _create_query_vector(self) -> Optional[List[Any]]:
        return AzureQueryResultSearchDefault._create_query_vector(self)

    def _create_search_query(self) -> str:
        return AzureQueryResultSearchSparse._create_search_query(self)

As I can see it is not integrating alpha value. If this is true, what could be workaround here for alpha parameter usage?

[dosubot[bot]]

Hey there, @madtur! I'm a bot here to help you with bugs, questions, and becoming a contributor. Let's squash this bug together! 🤖

The alpha parameter is indeed not affecting the results in hybrid search mode for Azure AIsearch llama-index integration because the implementation does not support it. The code explicitly logs a warning that "postgres hybrid search does not support alpha parameter," indicating that the alpha parameter is not utilized in the hybrid search logic ^[1].

As a workaround, you can use the KodaRetriever class from the LlamaIndex framework, which allows for flexible alpha tuning. This custom hybrid retriever dynamically determines the optimal alpha for a given query using an LLM to categorize the query. You can set custom alpha values for different categories, and the retriever will adjust the alpha parameter accordingly. Here is an example of how you might use the KodaRetriever:

retriever = KodaRetriever(
    index=vector_index,
    llm=Settings.llm,
    verbose=True,
    similarity_top_k=10
)
results = retriever.retrieve("What happened on Y2K?")

This approach provides a way to incorporate alpha tuning into your hybrid search logic, even though the default implementation does not support it ^[2][3].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Other}