Advanced Configuration
Advanced Configurations allow you to fine-tune retrieval and answer optimization settings for specific requirements.Accessing Advanced Configuration
- Navigate to Dev Tools > Advanced Configurations
- Search for the configuration you want to modify
- Select or provide the appropriate values
Available Configurations
Configuration Details
Re-Ranking
Re-ranking improves search quality by applying a secondary model to reorder initially retrieved chunks based on deeper semantic analysis. Search AI supports the following re-rankers.- Cross Encoder Re-Ranker - Uses the cross-encoder/ms-marco-MiniLM-L-6-v2 model. It’s lightweight, fast, and most suitable for English Language.
- BGE Re-Ranker - Uses BAAI/bge-reranker-v2-m3 model. It’s a lightweight re-ranking model and possesses multilingual capabilities.
- MixedBread Re-Ranker - Uses the mixedbread-ai/mxbai-rerank-large-v1 model, which is resource intensive to run and has a higher latency but delivers the highest accuracy and performance.
This feature doesn’t require training of the application.
Increasing the number of fields or chunks for reranking may lead to higher latency due to the added complexity in computation, data retrieval, and processing load. If you are using the generative model and re-ranking is also enabled, the overall latency includes the latency induced by the re-ranker and the LLM.
KNN Matching
KNN stands for K-Nearest Neighbors. In RAG applications, KNN matching retrieves the most relevant information from indexed data based on semantic similarity, i.e., finding the closest chunks matching a given query. There are two types of KNN matching methods:- Exact KNN - Finds the truly exact neighbors by comparing the query with every vector of the indexed content. This type of matching guarantees higher accuracy, and precision but can be computationally expensive and can affect performance.
- Approximate KNN - Uses different techniques to find the nearest neighbors quickly rather than comparing with every vector. This method may provide a close match rather than the best one, but it’s faster and more scalable for large datasets.
Enabling Exact KNN Match can introduce some latency and thereby increase the average Response Time.
Document Security
When a file is ingested into Search AI for indexing, a signed URL is automatically generated for the uploaded document on the server. These URLs are used as references or citations when search results or answers are derived from the corresponding document. The signed URL provides secure, temporary access to the document. It is valid for a single use or 5 minutes, whichever comes first. This ensures controlled access and prevents unauthorized sharing.This applies only to uploaded documents. References for data from connectors or web pages include a direct link to the corresponding web page or third-party application.
Encoding Model
An encoding model defines how text is split into tokens before being processed. This configuration defines the encoding model used for token counting across indexing, extraction, and runtime operations in Search AI. The tokenizer determines:- How many tokens a piece of text occupies — directly affecting whether content fits within model context windows
- Where content gets split — affecting the semantic coherence of chunks
- How batches are sized — affecting throughput and cost during indexing
By default, apps created in v11.23.1 or later use GPT-4o (
o200k_base) encoding models. For apps created before this version, update the Encoding Model and manually retrain the app.- GPT-4o —
o200k_base(Recommended) - GPT-4 / GPT-3.5 —
cl100k_base - GPT-2 —
gpt2(Legacy default model)
- Extraction and Document Processing — used when splitting content into chunks (text/CSV, HTML, Markdown, and AI Vision content; conversation chunking for Slack, Teams, and other message-based connectors)
- Enrichment and Vector Generation — used for embedding batch sizing and LLM-stage document truncation
- Answer Generation — used for token counting during search and response generation (token budget calculation, top-k chunk selection, response limits)
Enable Click Tracking
The Enable Click Tracking configuration allows tracking of user interactions with links in Search AI responses. When enabled, all chunk and citation URLs in answers are automatically wrapped with tracking links. Each click is recorded and mapped to the corresponding query and result, providing insights into user engagement. Upon click, the following data is captured: Upon click, the following data is captured:- Tracking is handled server-side, so no client-side instrumentation or code changes are required
- Works consistently across web, mobile, and API integrations
- Use the Answer Analytics API to download complete click analytics data
- Use the Click Analytics API for aggregated insights such as:
- Time-series trends
- Average click statistics (clicks per search, chunks clicked, total searches with clicks)
- Average click position
- Click position distribution
- Top queries with no clicks
Toolkit - Developer Utilities
The Toolkit provides SDKs and utilities for content processing, data extraction, performance evaluation, and custom connector development.Available Tools
Evaluation Tools
RAG Evaluator on GitHub
Integration Tools
Custom Connector SDK on GitHub
Extraction Utilities
Model Optimization
Fine-Tune Embedding Utility on GitHub