Ragtacts

Ask LLMs easily with Ragtacts!

Documents

• English

• 한국어

• ractacts cljdoc

Prerequisites

Install Clojure

Please install java and brew first to install Clojure. And install Clojure with the following command:

$ brew install clojure/tools/clojure

Clojure REPL with ragtacts

Create a deps.edn file and insert the following contents.

com.constacts/ragtacts {:mvn/version "0.4.1"}

Next, run the Clojure REPL with the following command. Since ragtacts uses OpenAI as the default LLM model, an OpenAI API key is required. Refer to the OpenAI documentation to prepare your key.

$ OPENAI_API_KEY=sk-xxxx clj
Clojure 1.11.3
user=> 

QuickStart

To use the Ragtacts library, you need to require the ractacts.core namespace.

(require '[ragtacts.core :refer :all])

Ask to an LLM (Large Language Model)

Put the question you want to ask in the argument of the ask function.

(ask "Hello!")
;; [{:user "Hello!"} {:ai "Hi there! How can I assist you today?"}]

The result of ask will be in the form of a question and answer. Each item in the result list is a map containing a role and content. The roles are :user and :ai. The last item with the LLM's answer will be the value associated with the :ai key.

The default model is OpenAI's gpt-4 but you can also ask questions to other models.

(-> "Hello!"
    (ask {:model "gpt-4-turbo"})
    last
    :ai)
;; "Hi there! How can I assist you today?"

You can create question templates using the prompt function. The templates follow the Python str.format template syntax.

(-> "Tell me a {adjective} joke about {content}."
    (prompt {:adjective "funny" :content "chickens"})
    ask
    last
    :ai)
;; "Sure, here's a classic one for you:\n\nWhy did the chicken go to the séance?\n\nTo ta..."

You can use prompts from the Langchain Hub.

(require '[ragtacts.prompt.langchain :as langchain])

(-> (langchain/hub "rlm/rag-prompt") 
    (prompt {:context "Ragtacts is an easy and powerful LLM library." 
             :question "What is Ragtacts?"})
    ask
    last
    :ai)
;; "Ragtacts is an easy and powerful LLM library."

If you use a model that supports multimodal inputs, you can also ask questions about images.

(->
   (ask (with-images "What are in these images? Is there any difference between them?"
     "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
     (io/input-stream "/tmp/sample.png")))
   last
   :ai)

When asking a question, if you provide previous conversation context, the response will be based on that conversation context.

(-> [{:system "You are a wondrous wizard of math."}
     {:user "2+2"}
     {:ai "4"}
     {:user "2+3"}
     {:ai "5"}
     {:user "What's the square of a triangle?"}]
    ask
    last
    :ai)
;; "The phrase \"square of a triangle\" is a ..."

Since the result of the ask function is conversation content, you can append the conversation to the result and call the ask function again to continue asking questions based on the previous conversation context.

(-> (ask "Hi I am Ragtacts")
    (conj "What is my name?")
    ask
    last
    :ai)
;; "You mentioned earlier that your name is Ragtacts. How can I help you today, Ragtacts?"

Invoke a Clojure function in natural language

You can call a Clojure function in natural language using the ask function. To let the LLM know what the function does, you need to include metadata in the function as follows.

(defn ^{:desc "Get the current weather in a given location"} get-current-weather
    [^{:type "string" :desc "The city, e.g. San Francisco"} location]
    (case (str/lower-case location)
      "tokyo" {:location "Tokyo" :temperature "10" :unit "fahrenheit"}
      "san francisco" {:location "San Francisco" :temperature "72" :unit "fahrenheit"}
      "paris" {:location "Paris" :temperature "22" :unit "fahrenheit"}
      {:location location :temperature "unknown"}))

(-> "What 's the weather like in San Francisco, Tokyo, and Paris?"
    (ask {:tools [#'get-current-weather]})
    last
    :ai)
;; "Here is the current weather in the requested cities:\n\n1. **San Francisco**: 72°F\n2. **Tokyo**: 
;;  10°F\n3. **Paris**: 22°F\n\nIt seems like the temperatures vary significantly across these cities!"  

In some cases, you need to use the result of calling a function as is. In such cases, you can use the :as key with the :values option to receive the result in the following form.

(-> "What 's the weather like in San Francisco, Tokyo, and Paris?"
    (ask {:tools [#'get-current-weather] :as :values})
    last
    :ai)

;; [{:get-current-weather {:location "San Francisco", :temperature "72", :unit "fahrenheit"}} 
;;  {:get-current-weather {:location "Tokyo", :temperature "10", :unit "fahrenheit"}} 
;;  {:get-current-weather {:location "Paris", :temperature "22", :unit "fahrenheit"}}]

The results are in a list because you can call the same function multiple times in one question. Each item contains the result value with the function name as the key. If multiple functions are included in :tools, the LLM can find and call the appropriate function, allowing you to know which function was called by its key.

Getting more accurate answers with a vector database.

A vector database stores data in vector format. Storing data as vectors allows for finding similar data. Suppose you ask an LLM about the contents of a book. The LLM may not be able to provide an accurate answer because it does not know the book's contents. However, if you include the book's contents in the LLM prompt, the LLM can reference it to give an accurate answer.

The problem is that the size of the prompt the LLM can handle is limited. Using a vector database can reduce the data to be included in the LLM prompt. By slicing the book's contents into smaller parts and storing them in a vector database, you can find several pieces of data most similar to the question and include them in the LLM prompt. This method is called RAG (Retrieval-Augmented Generation).

You can easily do RAG using Ragtacts. Let's first store and retrieve data in the vector database.

(let [db (vector-store)]
  (add db ["The new data outside of the LLM's original training data set is called external data."
            "What Is RAG?"
            "The next question may be—what if the external data becomes stale?"
            "Retrieval-Augmented Generation (RAG) is the process of optimizing the output of a large language model."
            "The next step is to perform a relevancy search."
            "Recursive summarization as Context Summarization techniques provide a condensed view of documents"])
  (search db "Tell me about RAG"))
;; ("What Is RAG?" "Retrieval-Augmented Generation (RAG) is the process of optimizing the output of a large language 
;;  model." "Recursive summarizat...)

The vector-store function creates an in-memory vector database. You can store documents in the vector database using the add function and retrieve the most similar documents using the search function, which by default fetches the 5 most similar documents in order. The number of documents retrieved can be changed using the top-k option value.

(let [db (vector-store)]
  (add db ["The new data outside of the LLM's original training data set is called external data."
            "What Is RAG?"
            "The next question may be—what if the external data becomes stale?"
            "Retrieval-Augmented Generation (RAG) is the process of optimizing the output of a large language model."
            "The next step is to perform a relevancy search."
            "Recursive summarization as Context Summarization techniques provide a condensed view of documents"])
  (search db "Tell me about RAG" {:top-k 2}))
;; ("What Is RAG?" "Retrieval-Augmented Generation (RAG) is the process of optimizing the output of a large language 
;;  model.")

You can include additional information along with the documents to be stored as vectors, and filter your search results using this additional information.

(let [db (vector-store)]
  (add db [{:text "What Is RAG?"
              :metadata {:topic "RAG"}}
            {:text "The next question may be—what if the external data becomes stale?"
              :metadata {:topic "Tutorial"}}
            {:text "The next step is to perform a relevancy search."
              :metadata {:topic "Tutorial"}}])
  (search db "Tell me about RAG" {:metadata {:topic "Tutorial"}}))
;; ("The next question may be—what if the external data becomes stale?" "The next step is to..")

The "What Is RAG?" was most similar to "Tell me about RAG", but since the search was filtered to only include documents with metadata where the topic is "Tutorial", "What Is RAG?" did not appear in the results.

If you add the {:raw? true} option to the search function, you can retrieve the stored vector values and metadata in the result.

(let [db (vector-store)]
  (add db [{:text "What Is RAG?"
              :metadata {:topic "RAG"}}
            {:text "The next question may be—what if the external data becomes stale?"
              :metadata {:topic "Tutorial"}}
            {:text "The next step is to perform a relevancy search."
              :metadata {:topic "Tutorial"}}])
  (search db "Tell me about RAG" {:metadata {:topic "Tutorial"}
                                  :raw? true}))
;; [{:text "The next step is to perform a relevancy search." 
;;   :vector [-0.002841026 0.015938155 ...]
;;   :metadata {:topic "Tutorial"}} ...]

You can extract text from web pages or documents (e.g., PDF, DOC, XLS, PPT) and store it in the vector database for searching.

(require '[ragtacts.loader.web :as web])

(let [db (vector-store)
      text (web/get-text "https://aws.amazon.com/what-is/retrieval-augmented-generation/")]
  (add db [text])
  (search db "What is RAG?"))

(require '[ragtacts.loader.doc :as doc])

(let [db (vector-store)
      text (doc/get-text "~/papers/RAPTOR.pdf")]
  (add db [text])
  (search db "What is RAPTOR?"))

As mentioned earlier, you can split the text and store it in the vector database. If the text passed to the add function is long, it will be split and stored in the vector database. The default value is 500 characters. The text is not cut exactly at 500 characters to avoid splitting in the middle of a sentence or word. You can change the character limit using the :splitter option in the vector-store function. You need to provide the :size and :overlap options. The :overlap option specifies the overlap size to ensure text is not cut off abruptly.

(let [db (vector-store {:splitter {:size 100 :overlap 10}})
      text (doc/get-text "~/papers/RAPTOR.pdf")]
  (add db [text])
  (search db "What is RAPTOR?"))

Now, let's ask the LLM based on the content in the vector database. We need to concatenate the retrieved content from the vector database into a string, incorporate it into an appropriate prompt, and then query the LLM. For the example, we will use the rlm/rag-prompt from the LangChain Hub as the prompt template.

(let [db (vector-store)
        text (web/get-text "https://aws.amazon.com/what-is/retrieval-augmented-generation/")
        rag-prompt (langchain/hub "rlm/rag-prompt")
        question "What is RAG?"]
    (add db [text])
    (-> (ask (prompt rag-prompt {:context (str/join "\n" (search db question))
                                 :question question}))
        last
        :ai))

Ragtacts has a watch function that can update the vector database with the changed content when the content on a web page or in a folder is updated. This function allows you to keep the data in the vector database synchronized with the changing data.

(def web-wather
  (web/watch {:url "https://aws.amazon.com/what-is/retrieval-augmented-generation/"
              :interval 1000}
              (fn [change-log]
                ;; {:type :create :text "..."}
                (println change-log))))

(web/stop-watch web-wather)

;; WIP
(def folder-wather
  (doc/watch {:path "~/papers"}
              (fn [change-log]
                (println change-log))))

(doc/stop-watch folder-wather)

Using Ragtacts RAG Playground

The examples folder contains a RAG Playground created with electric. Run the Playground with the following command and point your web browser to http://localhost:8080 in your web browser.

$ cd examples/playground
$ clj -A:dev -X dev/-main

Using Ragtacts as CLI

Ragtacts can also be used as a CLI. Download the ragtacts.jar file from the Releases page, and run it with Java to allow querying an LLM based on web pages or documents.

$ java -jar target/ragtacts-standalone.jar -p "What is RAG?" -d https://aws.amazon.com/what-is/retrieval-augmented-generation/
AI: RAG, or Retrieval-Augmented Generation, is a process that enhances the output of a large language model (LLM) by incorporating an information retrieval component. This component pulls relevant information from an external knowledge base and provides it to the LLM, enabling it to generate more accurate responses. This approach offers organizations better control over the generated text output and improves the overall quality of the responses. 

By using the chat mode, you can ask questions interactively.

$ java -jar target/ragtacts-standalone.jar -m chat -d https://aws.amazon.com/what-is/retrieval-augmented-generation/
Prompt: What is RAG?
AI: RAG, or Retrieval-Augmented Generation, is a process that optimizes the output of a large language model by first retrieving information from an external, authoritative knowledge base before generating a response. This allows the model to use both its training data and the new information to create more accurate and reliable answers. This approach gives organizations greater control over generated text and helps improve the quality of the responses. 
Prompt:

Using Ragtacts as API Server

You can also use Ragtacts as an API server. Enter the following command and then access http://localhost:3000. The API is compatible with the OpenAI Chat API.

$ java -jar target/ragtacts-standalone.jar -m server -d https://aws.amazon.com/what-is/retrieval-augmented-generation/

Contributing

Please read CONTRIBUTING.md before submitting a pull request.

License

Copyright © 2024 Constacts, Inc.

Distributed under the Eclipse Public License, the same as Clojure.