Contents x
    5. Adjust Reranker
      • PDF
      Contents

      5. Adjust Reranker

        • PDF
        Article summary

        ❓ What does the reranker do?

        • The reranker is like a second, intelligent filter that looks over it again after the first search and sorts the most important content better.

        • Especially in the case of similar terms, unclear questions or large documents, it improves the quality of the answer.

        Simply put:

        • The normal search asks: "Which text passages roughly match the question?"

        • The reranker also asks: "Which of these passages are really relevant and useful for the answer?"

        🛠️ How does it work exactly?

        First, the system searches for hits, e.g. with keyword or semantic searches. This results in 15 matching text sections (chunks) from your documents, for example.

        Then the reranker re-evaluates these 15 results by comparing them to the question asked – at the level of meaning

        • What is really helpful for the answer?

        • What is only vaguely or coincidentally similar?

        The reranker then rearranges the results: The best and most meaningful chunks go to the top – i.e. first in the AI response.

        ⚠️ Attention

        • Too large a context size in combination with the reranker weakens the speed. The search of the AI is carried out several times, so to speak.

        • If you can't get past the reranker and the speed is not sufficient, check out our verfügbaren KI Modelle .

        📝 Example

        The user asks: "Which laptop is suitable for external monitors with high image quality?"

        The product data/texts contain:

        • Model A has "HDMI 2.1" on page 3

        • Model B has "external ports" on page 5

        • Model C is "well suited for presentations" on page 8

        • The system simply lists all the places with "HDMI", "external", "image" and "presentation".

        • Model C (because of "presentation") could mistakenly end up at the top – although HDMI 2.1 would be much more relevant for Model A.

        • The reranker recognizes that "HDMI 2.1" fits directly to the question – namely "high image quality on external monitor".

        • Model A slides up. Model C is downgraded because "presentation" is too generic.

        🟢 Good if:

        • Your documents contain many similar terms or products

        • You want to avoid the AI taking "any hit"

        • You have complex questions, such as: "Which laptop is best for on the go AND has long battery life?"

        • You use semantic search and quality is more important than speed

        🔴 Not good if:

        • You only have very simple, clearly formulated questions

        • Your documents are very short or very clear

        • You need maximum speed (the reranker needs some computing time)

        Was this article helpful?
        What's Next