Major Update: AI Online Search Capability Now Available, Significantly Enhancing Dataset Hit Rate!
Aug 1, 2023
We are excited to announce the latest version, V0.3.12, which introduces two significant new features: further optimization of dataset retrieval strategies and the launch of "Chat" capability, enabling online interactions with AI. Below, we will provide a detailed overview of the new capabilities and features in this version.
Improved Q2Q Matching Mode in Datasets for Enhanced Data Retrieval
In private data applications, the quality of AI responses relies on the indexing strategy during the vectorization process. To continually enhance the quality and performance of AI applications in real-world scenarios, we have been exploring more accurate retrieval strategies. In our dataset's high-quality indexing approach, we have introduced the Q&A segmenting mode, which differs from the regular "Q2P" (Question-to-Paragraph) matching mode.
With the "Q2Q" (Question-to-Question) matching mode, when a user asks a question, the system identifies the most similar questions and returns the corresponding segments as answers. This approach is more precise because it directly matches the user's question, enabling a more accurate retrieval of the information the user truly needs.
For example, when presented with a data table containing various movie information such as titles, duration, synopsis, cast, etc., and asked about the main storyline of the movie "Spider-Man: Across the Spider-Verse," the Q&A segmenting mode feature, when enabled, allows the system to accurately identify dataset questions and answers, providing precise outputs.
"The film 'Spider-Man: Across the Spider-Verse' primarily focuses on the story of Miles Morales, a Brooklyn teen who finds himself becoming Spider-Man and having to team up with other Spider-Men from different dimensions to save the world."
Without the Q&A segmenting mode enabled, the system would have been unable to provide a relevant answer and might prompt the user to provide more details for a better response.
"I don't know, but I can help you find the answer. Can you provide some more details about what you are looking for?"
Furthermore, if the Q&A segmenting mode result does not meet expectations, users can manually modify or add segment ranges and content to better align the dataset's structure and improve response accuracy.
Chat: Online Content Search during Conversations
Dify's mission is to explore the boundaries of large models together with developers. Since the introduction of capabilities like Plugin, Code Interpreter, and Function Call, developers have moved beyond the traditional prompt-based interaction with AI. With "Chat," Dify takes a significant step towards becoming an intelligent agent platform. In Chat, we have launched a set of first-party plugins, including web browsing, Google search, and Wikipedia query.
Through Explore, users can access Chat and achieve various functionalities, such as:
Prompting AI to answer time and computation-related questions
Analyzing web page content as context for the conversation
Expanding on AI's reasoning process
Chat serves as a productivity tool for team users and replaces the recently discontinued online access capability of ChatGPT.
Moreover, Chat will be a collaborative exploration laboratory for Dify and developers to build large model capabilities. New capabilities such as autonomous agents, plugin development, multimodal features, etc., will be validated in Chat before being integrated into Dify's application orchestration functions. Chat currently supports all models with reasoning capabilities, such as GPT-3.5, GPT-4, Claude 2, and all plugins are developed and optimized by the Dify team.
Thank you for being part of Dify, and we look forward to your continued support and feedback as we continue to evolve and enhance our platform.
More New Features
App Orchestration Debugging Page: Prompt Suggestions and Variable Displays now support collapsing.
Enhanced Documentation for the Speech-to-Text Service API Application Interface.
Various Bug Fixes.