diff --git a/en/.gitbook/assets/image (1) (1) (1).png b/en/.gitbook/assets/image (1) (1) (1).png new file mode 100644 index 0000000..8c20eca Binary files /dev/null and b/en/.gitbook/assets/image (1) (1) (1).png differ diff --git a/en/.gitbook/assets/image (1) (1).png b/en/.gitbook/assets/image (1) (1).png index 8c20eca..d11476b 100644 Binary files a/en/.gitbook/assets/image (1) (1).png and b/en/.gitbook/assets/image (1) (1).png differ diff --git a/en/.gitbook/assets/image (1).png b/en/.gitbook/assets/image (1).png index d11476b..abbd931 100644 Binary files a/en/.gitbook/assets/image (1).png and b/en/.gitbook/assets/image (1).png differ diff --git a/en/.gitbook/assets/image (2) (1) (1) (1).png b/en/.gitbook/assets/image (2) (1) (1) (1).png new file mode 100644 index 0000000..4f9324b Binary files /dev/null and b/en/.gitbook/assets/image (2) (1) (1) (1).png differ diff --git a/en/.gitbook/assets/image (2) (1) (1).png b/en/.gitbook/assets/image (2) (1) (1).png index 4f9324b..6ea5c61 100644 Binary files a/en/.gitbook/assets/image (2) (1) (1).png and b/en/.gitbook/assets/image (2) (1) (1).png differ diff --git a/en/.gitbook/assets/image (2) (1).png b/en/.gitbook/assets/image (2) (1).png index 6ea5c61..f19b0c0 100644 Binary files a/en/.gitbook/assets/image (2) (1).png and b/en/.gitbook/assets/image (2) (1).png differ diff --git a/en/.gitbook/assets/image (2).png b/en/.gitbook/assets/image (2).png index f19b0c0..eddf580 100644 Binary files a/en/.gitbook/assets/image (2).png and b/en/.gitbook/assets/image (2).png differ diff --git a/en/.gitbook/assets/screenshot-20231119-174228.png b/en/.gitbook/assets/screenshot-20231119-174228.png new file mode 100644 index 0000000..ccca8f6 Binary files /dev/null and b/en/.gitbook/assets/screenshot-20231119-174228.png differ diff --git a/en/.gitbook/assets/screenshot-20231119-174610.png b/en/.gitbook/assets/screenshot-20231119-174610.png new file mode 100644 index 0000000..f8403b5 Binary files /dev/null and b/en/.gitbook/assets/screenshot-20231119-174610.png differ diff --git a/en/.gitbook/assets/screenshot-20231119-175216.png b/en/.gitbook/assets/screenshot-20231119-175216.png new file mode 100644 index 0000000..6add1b3 Binary files /dev/null and b/en/.gitbook/assets/screenshot-20231119-175216.png differ diff --git a/en/.gitbook/assets/screenshot-20231119-175958.png b/en/.gitbook/assets/screenshot-20231119-175958.png new file mode 100644 index 0000000..685b650 Binary files /dev/null and b/en/.gitbook/assets/screenshot-20231119-175958.png differ diff --git a/en/.gitbook/assets/screenshot-20231119-182704.png b/en/.gitbook/assets/screenshot-20231119-182704.png new file mode 100644 index 0000000..4cb8212 Binary files /dev/null and b/en/.gitbook/assets/screenshot-20231119-182704.png differ diff --git a/en/.gitbook/assets/screenshot-20231119-191016.png b/en/.gitbook/assets/screenshot-20231119-191016.png new file mode 100644 index 0000000..a980a0a Binary files /dev/null and b/en/.gitbook/assets/screenshot-20231119-191016.png differ diff --git a/en/.gitbook/assets/screenshot-20231119-191531.png b/en/.gitbook/assets/screenshot-20231119-191531.png new file mode 100644 index 0000000..939ca46 Binary files /dev/null and b/en/.gitbook/assets/screenshot-20231119-191531.png differ diff --git a/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_LgAOVtxy9kQ0B8e2qaQl_image.webp b/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_LgAOVtxy9kQ0B8e2qaQl_image.webp new file mode 100644 index 0000000..4cd673a Binary files /dev/null and b/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_LgAOVtxy9kQ0B8e2qaQl_image.webp differ diff --git a/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_ohgmBurknjsKmg53Z00U_image.webp b/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_ohgmBurknjsKmg53Z00U_image.webp new file mode 100644 index 0000000..519856c Binary files /dev/null and b/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_ohgmBurknjsKmg53Z00U_image.webp differ diff --git a/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_xfMNnsyD506TOoynHdgU_image.webp b/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_xfMNnsyD506TOoynHdgU_image.webp new file mode 100644 index 0000000..6af0c26 Binary files /dev/null and b/en/.gitbook/assets/spaces_CdDIVDY6AtAz028MFT4d_uploads_xfMNnsyD506TOoynHdgU_image.webp differ diff --git a/en/SUMMARY.md b/en/SUMMARY.md index 03473a2..d98642f 100644 --- a/en/SUMMARY.md +++ b/en/SUMMARY.md @@ -42,11 +42,15 @@ * [Expert Mode for Prompt Engineering](advanced/prompt-engineering/README.md) * [Prompt Template](advanced/prompt-engineering/prompt-template.md) +* [RAG (Retrieval Augmented Generation)](advanced/retrieval-augment/README.md) + * [Hybrid Search](advanced/retrieval-augment/hybrid-search.md) + * [Rerank](advanced/retrieval-augment/rerank.md) + * [Retrieval](advanced/retrieval-augment/retrieval.md) * [Datasets\&Index](advanced/datasets/README.md) * [Sync from Notion](advanced/datasets/sync-from-notion.md) * [Maintain Datasets Via Api](advanced/datasets/maintain-dataset-via-api.md) -* [Plugins](advanced/ai-plugins.md) -* [Based on WebApp Template](advanced/based-on-frontend-templates.md) +* [Plugins](advanced/ai-plugins/README.md) + * [Based on WebApp Template](advanced/ai-plugins/based-on-frontend-templates.md) * [Model Configuration](advanced/model-configuration/README.md) * [Hugging Face](advanced/model-configuration/hugging-face.md) * [Replicate](advanced/model-configuration/replicate.md) diff --git a/en/advanced/ai-plugins.md b/en/advanced/ai-plugins/README.md similarity index 100% rename from en/advanced/ai-plugins.md rename to en/advanced/ai-plugins/README.md diff --git a/en/advanced/based-on-frontend-templates.md b/en/advanced/ai-plugins/based-on-frontend-templates.md similarity index 100% rename from en/advanced/based-on-frontend-templates.md rename to en/advanced/ai-plugins/based-on-frontend-templates.md diff --git a/en/advanced/datasets/README.md b/en/advanced/datasets/README.md index c9441d3..8590d4d 100644 --- a/en/advanced/datasets/README.md +++ b/en/advanced/datasets/README.md @@ -80,7 +80,7 @@ Dify support customizing the segmented and cleaned text by adding, deleting, and

Edit

-

add

+

add

### Disabling and Archiving of Documents diff --git a/en/advanced/datasets/sync-from-notion.md b/en/advanced/datasets/sync-from-notion.md index 709f715..5b1f424 100644 --- a/en/advanced/datasets/sync-from-notion.md +++ b/en/advanced/datasets/sync-from-notion.md @@ -43,7 +43,7 @@ Click the " **New integration** " button, the type is Internal by default (canno Once the integration is created, you can update its settings as needed under the **Capabilities** tab and click the "**Show**" button under **Secrets** and then copy the Secrets. -
+
Copy it and back to the Dify source code , in the **.env** file configuration related environment variables, environment variables as follows: @@ -57,7 +57,7 @@ Copy it and back to the Dify source code , in the **.env** file configuration re To toggle the switch to public settings, you need to **fill in additional information in the Organization Information** form below, including your company name, website, and Retargeting URL, and click the "Submit" button. -
+
After your integration has been successfully made public in your [integration’s settings page](https://www.notion.so/my-integrations), you will be able to access the integration’s secrets in the Secrets tab. diff --git a/en/advanced/retrieval-augment/README.md b/en/advanced/retrieval-augment/README.md new file mode 100644 index 0000000..e6e1cb7 --- /dev/null +++ b/en/advanced/retrieval-augment/README.md @@ -0,0 +1,21 @@ +# RAG (Retrieval Augmented Generation) + +## The concept of the RAG + +The RAG architecture, with vector retrieval at its core, has become the leading technological framework for addressing two major challenges of large models: acquiring the latest external knowledge and mitigating issues of generating hallucinations. This architecture has been widely implemented in numerous practical application scenarios. + +Developers can utilize this technology to cost-effectively build AI-powered customer service bots, corporate knowledge bases, AI search engines, etc. These systems interact with various forms of organized knowledge through natural language input. A representative example of a RAG application is as follows: + +In the diagram below, when a user asks, "Who is the President of the United States?", the system doesn't directly relay the question to the large model for an answer. Instead, it first conducts a vector search in a knowledge base (like Wikipedia, as shown in the diagram) for the user's query. It finds relevant content through semantic similarity matching (for instance, "Biden is the current 46th President of the United States…"), and then provides the user's question along with the found knowledge to the large model. This enables the model to have sufficient and complete knowledge to answer the question, thereby yielding a more reliable response. + +

Basic Architecture of RAG

+ +## Why is this necessary? + +We can liken a large model to a super-expert, knowledgeable in various human domains. However, this expert has its limitations; for example, it doesn't know your personal situation, as such information is private and not publicly available on the internet, and therefore, it hasn't had the opportunity to learn it beforehand. + +When you want to hire this super-expert as your family financial advisor, you need to allow them to review your investment records, household expenses, and other relevant data before they can respond to your inquiries. This enables them to provide professional advice tailored to your personal circumstances. + +**This is what the RAG system does: it helps the large model temporarily acquire external knowledge it doesn't possess, allowing it to search for answers before responding to a question.** + +Based on this example, it's evident that the most critical aspect of the RAG system is the retrieval of external knowledge. The expert's ability to provide professional financial advice depends on accurately finding the necessary information. If the expert retrieves information unrelated to financial investments, like a family weight loss plan, even the most capable expert would be ineffective. diff --git a/en/advanced/retrieval-augment/hybrid-search.md b/en/advanced/retrieval-augment/hybrid-search.md new file mode 100644 index 0000000..1c11fa9 --- /dev/null +++ b/en/advanced/retrieval-augment/hybrid-search.md @@ -0,0 +1,80 @@ +# Hybrid Search + +## Why is Hybrid Search Necessary? + +The mainstream method in the RAG retrieval phase is Vector Search, which is based on semantic relevance matching. The technical principle involves initially dividing documents from external knowledge bases into semantically complete paragraphs or sentences, and then converting them through a process known as embedding into a series of numerical expressions (multidimensional vectors) that computers can understand. The user's query undergoes a similar conversion. + +Computers can detect subtle semantic correlations between user queries and sentences. For example, the semantic relevance between "a cat chases a mouse" and "a kitten hunting a mouse" is higher than between "a cat chases a mouse" and "I like to eat ham." After identifying the text with the highest relevance, the RAG system provides it as context alongside the user's query to the large model, aiding in answering the question. + +Besides complex semantic text searches, Vector Search has other advantages: + +* Understanding of similar semantics (e.g., mouse/mousetrap/cheese, Google/Bing/search engine) +* Multilingual comprehension (e.g., matching Chinese input with English content) +* Multimodal understanding (supports matching text, images, audio, and video) +* Fault tolerance (handles spelling mistakes, vague descriptions) + +However, Vector Search might underperform in certain scenarios, like: + +* Searching names of people or objects (e.g., Elon Musk, iPhone 15) +* Searching acronyms or short phrases (e.g., RAG, RLHF) +* Searching IDs (e.g., `gpt-3.5-turbo`, `titan-xlarge-v1.01`) + +These limitations are precisely where traditional keyword search excels, being adept at: + +* Precise matching (e.g., product names, personal names, product numbers) +* Matching a small number of characters (vector search performs poorly with few characters, but users often input just a few keywords) +* Matching low-frequency vocabulary (which often carries more significant meanings, like in “Do you want to go for a coffee with me?”, words like “drink” and “coffee” carry more weight than “you”, “want”, “me”) + +In most text search scenarios, it's crucial to ensure that the most relevant results appear in the candidates. Vector and keyword searches each have their strengths in the search domain. Hybrid Search combines the advantages of both techniques while compensating for their respective shortcomings. + +In Hybrid Search, vector and keyword indices are pre-established in the database. Upon user query input, the system searches for the most relevant text in documents using both search methods. + +

Hybrid Search

+ +"Hybrid Search" doesn't have a definitive definition; this article exemplifies it as a combination of Vector Search and Keyword Search. However, the term can also apply to other combinations of search algorithms. For instance, we could combine knowledge graph technology, used for retrieving entity relationships, with Vector Search. + +Different search systems each excel at uncovering various subtle connections within texts (paragraphs, sentences, words), including precise relationships, semantic relationships, thematic relationships, structural relationships, entity relationships, temporal relationships, and event relationships. It's safe to say that no single search mode is suitable for all scenarios. **Hybrid Search, by integrating multiple search systems, achieves a complementarity among various search technologies.** + +## Vector Search + +Definition: Vector Search involves generating query embeddings and then searching for text segments that most closely match these embeddings in terms of vector representation. + +

Settings for Vector Search

+ +**TopK:** This setting is used to filter text segments that have the highest similarity to the user's query. The system also dynamically adjusts the number of segments based on the context window size of the selected model. The default value for this setting is 3. + +**Score Threshold:** This setting is used to establish a similarity threshold for the selection of text segments. It means that only text segments exceeding the set score are recalled. By default, this setting is turned off, meaning that the system does not filter the similarity values of the recalled text segments. When activated, the default value is set to 0.5. + +**Rerank Model:** After configuring the Rerank model's API key on the "Model Provider" page, you can enable the "Rerank Model" in the search settings. The system then performs a semantic re-ranking of the document results that have been recalled after semantic search, optimizing the order of these results. Once the Rerank model is set up, the TopK and Score threshold settings are only effective in the Rerank step. + +## Full-Text Search + +Definition: Full-Text Search involves indexing all the words in a document, enabling users to query any term and retrieve text segments that contain these terms. + +

Settings for Full-Text Search

+ +**TopK:** This setting is utilized to select text segments that most closely match the user's query in terms of similarity. The system also dynamically adjusts the number of segments based on the context window size of the chosen model. The default value for TopK is set at 3. + +**Rerank Model:** After configuring the API key for the Rerank model on the "Model Provider" page, you can activate the "Rerank Model" in the search settings. The system will then perform a semantic re-ranking of the document results retrieved through full-text search, optimizing the order of these results. Once the Rerank model is configured, the TopK and any Score threshold settings will only be effective during the Rerank step. + +## Hybrid Search + +Hybrid Search operates by concurrently executing Full-Text Search and Vector Search. It then applies a re-ranking step to choose the best results that match the user's query from both types of search results. To effectively use this feature, it is necessary to configure the Rerank Model API. + +

Settings for Hybrid Search

+ +**TopK:** This setting is used for filtering text segments that have the highest similarity to the user's query. The system will dynamically adjust the number of segments based on the context window size of the model in use. The default value for TopK is set at 3. + +**Rerank Model:** After configuring the Rerank model's API key on the "Model Supplier" page, you can enable the "Rerank Model" in the search settings. The system will perform a semantic re-ranking of the document results retrieved through hybrid search, thereby optimizing the order of these results. Once the Rerank model is set up, the TopK and any Score threshold settings are only applicable during the Rerank step. + +## Setting the Search Mode When Creating a Dataset + +To set the search mode when creating a dataset, navigate to the "Datasets -> Create Dataset" page. There, you can configure different search modes in the retrieval settings section. + +

Setting the Search Mode When Creating a Dataset

+ +## Modifying the Search Mode in Prompt Engineering + +You can modify the search mode during application creation by navigating to the "Prompt Engineering -> Context -> Select Dataset -> Settings" page. This allows for adjustments to different search modes within the prompt arrangement phase. + +

Modifying the Search Mode in Prompt Engineering

diff --git a/en/advanced/retrieval-augment/rerank.md b/en/advanced/retrieval-augment/rerank.md new file mode 100644 index 0000000..37eb07d --- /dev/null +++ b/en/advanced/retrieval-augment/rerank.md @@ -0,0 +1,41 @@ +# Rerank + +## Why is Rerank Necessary? + +Hybrid Search combines the advantages of various search technologies to achieve better recall results. However, results from different search modes need to be merged and normalized (converting data into a uniform standard range or distribution for better comparison, analysis, and processing) before being collectively provided to the large model. This necessitates the introduction of a scoring system: Rerank Model. + +The Rerank Model works by reordering the list of candidate documents based on their semantic match with the user's question, thus improving the results of semantic sorting. It does this by calculating a relevance score between the user's question and each candidate document, returning a list of documents sorted by relevance from high to low. Common Rerank models include Cohere rerank, bge-reranker, and others. + +

Hybrid Search + Rerank

+ +In most cases, there is an initial search before rerank because calculating the relevance score between a query and millions of documents is highly inefficient. **Therefore, rerank is typically placed at the end of the search process, making it very suitable for merging and sorting results from different search systems.** + +However, rerank is not only applicable to merging results from different search systems. Even in a single search mode, introducing a rerank step can effectively improve the recall of documents, such as adding semantic rerank after keyword search. + +In practice, apart from normalizing results from multiple queries, we usually limit the number of text segments passed to the large model before providing the relevant text segments (i.e., TopK, which can be set in the rerank model parameters). This is done because the input window of the large model has size limitations (generally 4K, 8K, 16K, 128K Token counts), and you need to select an appropriate segmentation strategy and TopK value based on the size limitation of the chosen model's input window. + +It should be noted that even if the model's context window is sufficiently large, too many recalled segments may introduce content with lower relevance, thus degrading the quality of the answer. Therefore, the TopK parameter for rerank is not necessarily better when larger. + +Rerank is not a substitute for search technology but an auxiliary tool to enhance existing search systems. **Its greatest advantage is that it not only offers a simple and low-complexity method to improve search results but also allows users to integrate semantic relevance into existing search systems without the need for significant infrastructure modifications.** + +## How to Obtain the Cohere Rerank Model? + +Visit [https://cohere.com/rerank](https://cohere.com/rerank), register on the page, and apply for usage rights for the Rerank model to obtain the API key. + +## Setting the Rerank Model in Dataset Search Mode + +Access the Rerank settings by navigating to “Datasets -> Create Dataset -> Retrieval Settings”. Besides setting Rerank during dataset creation, you can also modify the Rerank configuration in the settings of an already created dataset, and change the Rerank configuration in the dataset recall mode settings of application arrangement. + +

Setting the Rerank Model in Dataset Search Mode

+ +**TopK:** Used to set the number of relevant documents returned after Rerank. + +**Score Threshold:** Used to set the minimum score for relevant documents to be returned after Rerank. After setting the Rerank model, the TopK and Score threshold settings are only effective in the Rerank step. + +## Setting the Rerank Model in Multi-path retrieval + + Recall Mode Enable the Rerank model by setting it to Multi-path retrieval mode in the “Prompt Engineering -> Context -> Settings” page. + +Explanation of Multi-path retrieval Mode: 🔗 + +

Setting the Rerank Model in Multi-path retrieval

diff --git a/en/advanced/retrieval-augment/retrieval.md b/en/advanced/retrieval-augment/retrieval.md new file mode 100644 index 0000000..788c0e7 --- /dev/null +++ b/en/advanced/retrieval-augment/retrieval.md @@ -0,0 +1,33 @@ +# Retrieval + +When users build knowledge base Q\&A AI applications, if multiple datasets are associated within the application, Dify supports two retrieval modes: N-to-1 retrieval and Multi-path retrieval. + +

Retrieval Settings

+ +## Retrieval **Settings** + +### **N-to-1 Retrieval** + +Based on user intent and dataset description, the Agent independently determines and selects the most matching single dataset for querying relevant text. This mode is suitable for applications with distinct datasets and a smaller number of datasets. N-to-1 retrieval relies on the model's inference capability to choose the most relevant dataset based on user intent. When inferring the dataset, the dataset serves as a tool for the Agent, chosen through intent inference; the tool description is essentially the dataset description. + +When users upload datasets, the system automatically creates a summary description of each dataset. To achieve the best retrieval results in this mode, you can view the system-generated summary description under “Datasets -> Settings -> Dataset Description” and check if this content clearly summarizes the dataset's content. + +Here is the technical flowchart for N-to-1 retrieval: + +

N-to-1 Retrieval

+ +Therefore, this mode's recall effectiveness can be impacted when there are too many datasets or when the dataset descriptions lack sufficient distinction. This mode is more suitable for applications with fewer datasets. + +Tip: OpenAI Function Call already supports multiple tool calls, and Dify plans to upgrade this mode to "N-to-M retrieval" in future versions. + +### Multi-path Retrieval + +Based on user intent, this mode matches all datasets simultaneously, queries relevant text segments from multiple datasets, and after a re-ranking step, selects the best results matching the user's question from the multi-path query results. Configuring the Rerank model API is required. In Multi-path retrieval mode, the search engine retrieves text content related to the user's query from all datasets associated with the application, merges the results from multi-path recall, and re-ranks the retrieved documents semantically using the Rerank model. + +In Multi-path retrieval mode, configuring the Rerank model is necessary. How to configure the Rerank model: 🔗 + +Here is the technical flowchart for Multi-path retrieval: + +

Multi-path retrieval

+ +As Multi-path retrieval does not rely on the model's inferencing capability or dataset descriptions, this mode can achieve higher quality recall results in multi-dataset searches. Additionally, incorporating the Rerank step can effectively improve document recall. Therefore, when creating a knowledge base Q\&A application associated with multiple datasets, we recommend configuring the retrieval mode as Multi-path retrieval. diff --git a/en/application/prompt-engineering/conversation-application.md b/en/application/prompt-engineering/conversation-application.md index c86501f..d269687 100644 --- a/en/application/prompt-engineering/conversation-application.md +++ b/en/application/prompt-engineering/conversation-application.md @@ -22,7 +22,7 @@ Click the "Create Application" button on the homepage to create an application. After the application is successfully created, it will automatically redirect to the application overview page. Click on the left-hand menu: “**Prompt Eng.**” to compose the application. -
+
**2.1 Fill in Prompts**