Overview
Synthesis AI Search uses retrieval-augmented generation (RAG) to provide answers based on your firm’s content. This means that every response is generated dynamically, rather than being pulled from a static database of pre-written answers. While responses should remain consistent in their factual accuracy, they may vary slightly in wording, structure, or emphasis each time you ask the same question.
This article explains why that happens and what you can expect when using AI Search.
Key Factors That Affect Responses
1. Your Content is Constantly Changing
Synthesis AI Search retrieves information from your firm's intranet pages, documents, videos, posts, and other sources. These knowledge sources are dynamic:
- New content is added (e.g., a recent policy update or training video).
- Older content is archived or updated, affecting what is considered relevant.
- Freshness matters—the system prioritizes newer content when appropriate.
Even if you submit the same query multiple times, the underlying content set may have changed.
2. Search Results Are Dynamically Retrieved
Before generating a response, AI Search finds the most relevant pieces of content by using a hybrid search approach that combines:
- Keyword Search – Finds content based on exact matches to terms in your query.
- Vector (Semantic) Search – Understands the meaning of your query and finds content based on contextual similarity, even if exact words aren’t used.
This retrieval process ensures that AI Search is always working with the most relevant information, but relevance scores can shift slightly over time, affecting which content is selected.
3. Generative AI Uses a Probabilistic Approach
Once the system retrieves the most relevant content chunks, it sends them to a large language model (LLM) to generate a summarized response.
LLMs do not store fixed answers; instead, they generate responses on the fly based on probabilities. This means:
- The facts remain the same, but the wording, structure, and emphasis may differ.
- Similar to how a human explains the same concept slightly differently each time, AI-generated summaries will naturally vary.
This is a fundamental property of how generative AI works—it does not retrieve memorized responses but constructs answers dynamically based on retrieved information.
4. Context and Query Interpretation Can Vary
AI Search also interprets your query before retrieving information. While most simple queries return consistent results, more complex queries might be understood slightly differently across multiple attempts, leading to subtle differences in responses.
What This Means for You
- Responses should remain factually accurate, even if wording or structure changes.
- If you notice major discrepancies, check if new content has been added or if older information has been removed.
- The system is designed to prioritize relevance, accuracy, and freshness, but if you ever see an unexpected response, you can flag it for review.
AI-generated responses, like human explanations, may evolve slightly over time, but they will always be based on the best available information.
If you have further questions, reach out to our support team at support@knowledge-architecture.com.