<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Vassilios Antonopoulos on Medium]]></title>
        <description><![CDATA[Stories by Vassilios Antonopoulos on Medium]]></description>
        <link>https://medium.com/@vantonop?source=rss-ffa5cd2a8118------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*OlBEy7w1RXT3M-5k.</url>
            <title>Stories by Vassilios Antonopoulos on Medium</title>
            <link>https://medium.com/@vantonop?source=rss-ffa5cd2a8118------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 20:22:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@vantonop/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Are Llama3 and Mixtral open models ready for RAG in Greek?]]></title>
            <link>https://medium.com/11tensors/are-llama3-and-mixtral-open-models-ready-for-rag-in-greek-7cc1ab7579ed?source=rss-ffa5cd2a8118------2</link>
            <guid isPermaLink="false">https://medium.com/p/7cc1ab7579ed</guid>
            <category><![CDATA[llama-3]]></category>
            <category><![CDATA[mistral]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[mixtral-8x22b]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <dc:creator><![CDATA[Vassilios Antonopoulos]]></dc:creator>
            <pubDate>Tue, 30 Apr 2024 06:38:31 GMT</pubDate>
            <atom:updated>2024-04-30T06:38:31.903Z</atom:updated>
            <content:encoded><![CDATA[<p>Open source models have been shaping the AI field in recent weeks. Mistral released a new MoE 8x22B model and a few days ago Meta introduced the first models of its Llama3 series with 8B and 70B parameters (a larger one 400B is expected later this year). These models climbed near the top of all the existing leaderboards (Huggingface, LMSYS) proving that they are extremely powerful for a large variety of applications. Although they seem to better support more languages, of course Greek is not among them. So, at this point in time, are we still left with closed systems like those of Open AI and Cohere for applications dealing with Greek textual data?</p><p>Let’s see where we are.</p><p>We have interacted with Llama3–70B and Mixtral MoE 8x22B. You could clearly understand that the models are quite capable of conversing in Greek and following instructions to handle Greek content in several scenarios. It is obvious that they can understand very well the meaning and concepts of the Greek text. Mixtral MoE 8x7B was also quite capable, although you could notice some grammatical or syntactical errors in its text generations.</p><p>One of the most common AI components is RAG (Retrieval Augmented Generation). Especially if you only need AI on your data. RAG is in the heart of most AI systems today. In 11tensors we have developed end-to-end solutions incorporating RAG architectures for systems such as:</p><ol><li>Conversational agents (chatbots)</li><li>Question — Answering based on own data (external Knowledge Bases)</li><li>Search engines</li></ol><p>Multiple LLMs in multiple roles can be used in the above systems. We have focused on the generator role of the RAG architecture, the LLM that combines the retrieved text segments and forms the final response to the user, and made an evaluation, which is quite interesting.</p><p>So for this task, we started by selecting a similar benchmark. We have used the Greek version of the belebele dataset, which is a reading comprehension task containing multiple choice questions on a given text excerpt (<a href="https://arxiv.org/abs/2308.16884">https://arxiv.org/abs/2308.16884</a>). Recently, Institute for Language and Speech Processing (ILSP/Athena R.C.) has released a fine-tuned version of Mistral-7B-v0.1, called Meltemi, pretrained and fine-tuned in Greek (<a href="https://huggingface.co/ilsp/Meltemi-7B-v1">https://huggingface.co/ilsp/Meltemi-7B-v1</a>). Although the Instruct version of the model lacks performance, compared to what we are used to from the existing top LLMs and needs further improvement in user alignment and instruction following, both its versions are very good in generating correct Greek text. This model is therefore a benchmark for Greek. We have successfully replicated the published 0.6367 score of the base model in the belebele ell task using EleutherAI LM Evaluation Harness and compared it to the Llama3 8B base and Instruct models. The results are shown in the table below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/348/1*Q02w2zazHJNb5NqrZBPRuw.png" /></figure><p>It is interesting to note that the Llama3 small model of 8B parameters clearly outperforms the Greek model in this Greek benchmark. Llama3 seems to have a very high comprehension level when handling Greek texts. This observation brought us to the next step. Compare the models on a RAG task.</p><p>11tensors possesses the <strong>largest available online Greek textual corpus</strong>, crawled from public sources between 2009–2021 and cleaned for media analysis and information extraction. We use this corpus to synthetically create datasets and fine-tune LLMs according to our clients’ needs. We have previously created a proprietary RAG-specific dataset that contains <strong>triplets of Questions-Context-Answers</strong>. We use this dataset for fine-tuning LLMs and in this case, we have fine-tuned <strong>Meltemi-7B-Instruct-v1</strong> and created the model <strong>meltemi-11tensors-RAG</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*N3S19IbC5pEyHR20CUcv7A.png" /></figure><p>We needed a similar dataset for our RAG evaluation, so we created a new one by selecting a random set of 50 documents and creating 5 question-answer pairs for each one of them, resulting in a <strong>final set of 250 triplets</strong>. Starting with these questions, we implemented a typical RAG workflow. Documents were initially chunked, embedded with a Sentence Transformer model and imported into a Vector database. Then, starting with the set of 250 questions, we retrieved the 3 most similar chunks from the database to use as common context for all the LLMs we wanted to evaluate. We didn’t optimize this procedure on purpose, so that all models are evaluated under common conditions and rather difficult cases.</p><p>The models we evaluated were the following:</p><ul><li>Meltemi-7B-Instruct-v1</li><li>Meltemi-11tensors-RAG</li><li>Llama3–8B-Instruct</li><li>Llama3–70B-Instruct (served by <a href="https://wow.groq.com/">Groq</a> — btw… insane token/sec speed, just insane!)</li><li>Mistral Mixtral MoE 8x22B (served by <a href="https://mistral.ai/">Mistral AI</a> API)</li></ul><p>For the evaluation we followed the <strong>LLM-as-a-Judge</strong> approach and have used <strong>Open AI’s GPT3.5 model</strong>. We have selected the 2 metrics that measure the performance of RAG’s last component that we are interested in: Answer correctness (compared to ground truth) and relevancy (compared to query and context) using Prometheus model templates as discussed <a href="https://huggingface.co/learn/cookbook/en/rag_evaluation">here</a>. The results are presented below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/471/1*mgZ-cN_alapx_8P5GSIn9A.png" /></figure><p>There are some conclusions and observations that we can make:</p><ul><li>11tensors’ fine-tuned model, although significantly smaller, performs on par with the larger models. Not a surprise though… This model is fine-tuned for this specific task. This is the beauty of adapting and using smaller models for specific tasks.</li><li>The previous observation also indicates the importance and the potential of Meltemi LLM. It seems that it can be a basis to build very efficient solutions in various applications dealing with Greek text.</li><li>Both the Llama3–70B and Mixtral models confirm that they can understand Greek language very well and also successfully generate syntactically and grammatically correct Greek text. These models are very powerful in reasoning, in instruction following and the fact that they perform so well in this task signifies that they can be further used in a wide range of applications involving Greek text.</li><li>The Llama3–8B-Instruct model, while performing very well on the reading comprehension task, struggles to generate correct Greek text efficiently on this taskand its score seems low. However, we expect it to perform much better if it is fine-tuned on the task (see <a href="https://medium.com/11tensors/a-greek-llm-after-fine-tuning-llama2-7b-678a8eca1ab3">here</a> our post for Llama2), although the extent of its general applicability compared to the larger models is questionable.</li></ul><p>Finally, to return to our original question that led to this analysis… well… actually yes! Although Greek is underrepresented in their training datasets, compared to the widely spoken languages, the models are still very capable and demonstrate a robust performance. Moreover, they offer the flexibility for further fine-tuning to tailor them specifically to our unique use cases.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7cc1ab7579ed" width="1" height="1" alt=""><hr><p><a href="https://medium.com/11tensors/are-llama3-and-mixtral-open-models-ready-for-rag-in-greek-7cc1ab7579ed">Are Llama3 and Mixtral open models ready for RAG in Greek?</a> was originally published in <a href="https://medium.com/11tensors">11tensors</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Connect an AI agent with your API: Intel Neural-Chat 7b LLM can replace Open AI Function Calling]]></title>
            <link>https://medium.com/11tensors/connect-an-ai-agent-with-your-api-intel-neural-chat-7b-llm-can-replace-open-ai-function-calling-242d771e7c79?source=rss-ffa5cd2a8118------2</link>
            <guid isPermaLink="false">https://medium.com/p/242d771e7c79</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[intel]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Vassilios Antonopoulos]]></dc:creator>
            <pubDate>Wed, 06 Dec 2023 13:26:56 GMT</pubDate>
            <atom:updated>2023-12-06T13:26:56.438Z</atom:updated>
            <content:encoded><![CDATA[<p>Various pretrained Large Language Models (LLMs) today come with notable reasoning capabilities, enabling them to break down intricate issues into simpler steps, offering solutions, actions, and evaluations at each step. By leveraging the sophisticated natural language understanding capabilities of LLMs, <strong>AI agents</strong> can interpret and generate human-like text, enabling them to comprehend complex queries, engage in dialogue, and perform various knowledge-intensive and language-related tasks with a high degree of precision and context sensitivity.</p><p>AI Agents’ great potential and power come with their capability to <strong>use external tools</strong>, such as a company’s API, a search engine, a RAG component for getting answers using internal private data or another specialized LLM (e.g. code generator).</p><p>The most frequent way to connect an AI agent to an external tool is by <strong>connecting it to an API</strong>. This means that the agent backed by the LLM should be able to understand when a request needs information provided by the API, then pause the processing and call the API with the correct set of parameters, get the response, incorporate it to the rest of information and continue the processing to reach the final response to the initial request. These requests can come from users (conversational agents), other software components or even other LLMs.</p><p>The <strong>challenge</strong> here for the agent is to not only understand that a call to an external API is needed and interrupt the processing to do it, but also be able to select the appropriate method from the API and of course use the correct set of parameters to do the call. During the pre-LLM period, we were building agents using a Natural Language Understanding (NLU) component. This component was either a custom trained intent classifier combined with an Entity Extraction module, or a similar component offered by third party NLUs like Dialogflow, wit.ai and rasa.ai trained on sample utterances/samples of our application. In a similar way, an LLM could be used now as an intent classifier and an entity extraction module in subsequent calls by leveraging appropriate prompts. So, one could use several LLM calls to mimic the previous flow and step-by-step try to determine if an API should be called, which API and method is the one to call and with which parameters. But that doesn’t seem optimal now. Current LLMs should be able to do this more efficiently and in a more automated manner. Is this the case?</p><p><strong>Open AI</strong> offers two models, the latest be gpt-3.5-turbo-1106 and gpt-4–1106-preview, which have been trained to both <strong>detect when a function should be called</strong> (depending on the model input) and then to <strong>respond with JSON</strong> that adheres to the function signature. This is called <strong>Function Calling</strong> and it’s the proposed way to connect an agent with an external API.</p><p>Except from detecting the API function to call, what is really critical for the LLM in this case is to be able to reliably return structured data back, in a standard format, otherwise we won’t be able to parse the response and determine when, which and how to call an API method. The specifications of the API methods are passed to the model via the prompt.</p><p><strong>11tensors</strong> has developed an autonomous <strong>AI fashion agent</strong> with vision and text understanding capabilities that learns the customer’s style and helps them shop. The agent connects to a <strong>fashion marketplace via its API</strong>. The initial implementation was done based on <strong>Open AI’s models and Function Calling</strong>. Both models proved to be very powerful and robust in this use case. They were consistent in returning JSON when a function call was needed, precise on the selection of the function and its parameters and intelligent in inferring metadata from the conversation history in order to effectively act even in rather difficult requests, such as <em>“I think I prefer the second dress you showed to me earlier, please provide me with some other similar options”</em>. Here it needs to go back in the conversation history, find the dresses request-response, point to the second product of the list, get its ID and then call the corresponding function.</p><p>The agent performed very good in all the use cases. But the tokens were consumed very fast :-) ! We have reached a solution; however, it exceeded the anticipated budget. So, we searched for an alternative one. Can an open-source LLM achieve the performance of Open AI’s models and offer the same Function Calling feature at a comparable performance?</p><p>We have tested <strong>Llama2–7b-chat</strong> and <strong>gorilla-openfunctions-v1</strong> models. The second one is fine-tuned for the specific usage (API calling). They both had a decent performance. It proved really hard to make Llama2 return structured responses, but the gorilla model was quite consistent at this. Llama2 was better in inference, gorilla had problem with the tricky cases, but both of them were lacking a lot the performance we had with Open AI. So, we were a bit disappointed.</p><p>It was then that <strong>Intel</strong> announced its <strong>neural-chat-7b-v3–1</strong> LLM, a fine-tune of <strong>mistralai/Mistral-7B-v0.1</strong>, which had scored extremely well in several reasoning benchmarks. Mistral-7b stands out as the top-performing model with fewer than 30 billion parameters, surpassing numerous larger models across various benchmarks. So, this fine-tune seemed very interesting and promising, and we decided to give it a try.</p><p>Neural-chat model is not trained in responding with JSON, nor in function calling. In order to reach a point where we get the same functionality as Open AI Function Calling, we need to think a bit differently. Our approach to achieve this was to introduce a concurrent <strong>2-mode operation</strong> of the LLM:</p><ul><li>Function Calling mode (we call it “tool”)</li><li>Conversation mode (we call it “chat”)</li></ul><p>We prompt differently the two modes. User messages are firstly checked for possible “tool” need (API call) using the Function Calling mode. Function description in the prompt and the motivation to try this method came from a very insightful post of Aaron Blondeau (<a href="https://dev.to/aaronblondeau/creating-a-llama-2-agent-6fj">https://dev.to/aaronblondeau/creating-a-llama-2-agent-6fj</a>).</p><p>Inside the prompt, we are providing the description of the API, few-shot examples, the history of the conversation, the message from the user and we repeat the request for JSON response in case a function should be called.</p><pre>tool_prompt = &#39;&#39;&#39;<br>      ### System:<br>      Assistant is a expert JSON builder designed to assist the users of a women&#39;s fashion marketplace.<br>      Assistant is able to trigger actions for User by responding with JSON strings that contain &quot;action&quot; and &quot;action_input&quot; parameters.<br><br>      Actions available to Assistant are:<br><br>      - &quot;recommend_clothes&quot;: Useful for when Assistant is asked to provide clothing recommendations, which can be filtered on price, designer brand, season and clothing category.<br>        - To use the recommend_clothes tool, Assistant should respond like so:<br>          {{&quot;action&quot;: &quot;recommend_clothes&quot;, &quot;action_input&quot;: {{&quot;category&quot;: &quot;Dresses&quot;, &quot;season&quot;: &quot;Winter&quot;, &quot;min_price&quot; : 20, &quot;max_price&quot;: 550, &quot;designer&quot;: &quot;mpatmos&quot;}}}}<br>      - &quot;get_product_details&quot;: Useful for when Assistant is asked to retrieve all metadata details of a specific clothing item product.<br>        - To use the get_product_details tool, Assistant should respond like so:<br>          {{&quot;action&quot;: &quot;get_product_details&quot;, &quot;action_input&quot;: {{&quot;productID&quot;: 55783296}}}}<br>      - &quot;get_similar_products&quot;: Useful for when Assistant is asked to find clothes similar to the one mentioned.<br>        - To use the get_similar_products tool, Assistant should respond like so:<br>          {{&quot;action&quot;: &quot;get_similar_products&quot;, &quot;action_input&quot;: {{&quot;productID&quot;: 12748937}}}}<br>      - &quot;get_designers_list&quot;: Useful for when Assistant is asked to provide a list of the designers or brands contained in the marketplace.<br>        - To use the get_designers_list tool, Assistant should respond like so:<br>          {{&quot;action&quot;: &quot;get_designers_list&quot;, &quot;action_input&quot;: &quot;all&quot;}}<br>      - &quot;get_designer_information&quot;: Useful for when Assistant is asked to provide information about a specific designer or brand.<br>        - To use the get_designer_information tool, Assistant should respond like so:<br>          {{&quot;action&quot;: &quot;get_designer_information&quot;, &quot;action_input&quot;: {{&quot;name&quot;: &quot;mpatmos&quot;}}}}<br><br>      Here are some past examples of Assistant responding to the User:<br><br>      User: Hey how are you today?<br>      Assistant: I&#39;m good thanks, how are you?<br>      User: Please, show me some warm dresses you have for the winter.<br>      Assistant: {{&quot;action&quot;: &quot;recommend_clothes&quot;, &quot;action_input&quot;: {{&quot;category&quot;: &quot;Dresses&quot;, &quot;season&quot;: &quot;Winter&quot;, &quot;min_price&quot; : 0, &quot;max_price&quot;: 550, &quot;designer&quot;: &quot;&quot;}}}}<br>      User: Interesting! Give me more details on the Abracatabra dress please.<br>      Assistant: {{&quot;action&quot;: &quot;get_product_details&quot;, &quot;action_input&quot;: {{&quot;productID&quot;: 43731284}}}}<br>      User: Very nice! Can you get me some others similar to the Trijichta pants?<br>      Assistant: {{&quot;action&quot;: &quot;get_similar_products&quot;, &quot;action_input&quot;: {{&quot;productID&quot;: 386786777}}}}<br>      User: Who are the designers and the brands you collaborate with?.<br>      Assistant: {{&quot;action&quot;: &quot;get_designers_list&quot;, &quot;action_input&quot;: &quot;all&quot;}}<br>      User: Please give me more information for the happyfrenchgang brand.<br>      Assistant: {{&quot;action&quot;: &quot;get_designer_information&quot;, &quot;action_input&quot;: {{&quot;name&quot;: &quot;happyfrenchgang&quot;}}}}<br>      User: Thanks, Bye!<br>      Assistant: See you later.<br>      <br>      The history of the current conversation between Assistant and User is:<br>      {0}<br>      <br>      ### User:<br>      {1}<br><br>      Respond with JSON string that contains &quot;action&quot; and &quot;action_input&quot; parameters if a tool is needed.<br>      ### Assistant: &#39;&#39;&#39;</pre><p>If the response from the “tool” LLM contains the requested JSON, so that it specifies that a function call should be made, we are doing it and then the response from the marketplace API (in JSON too) is given as Context to the “chat” LLM. Otherwise, the user request is forwarded as-is to the “chat” mode LLM, which is responsible for the conversation part and the “tool” response is totally ignored.</p><pre>def check_tool(command):<br>    <br>    # Put user command into &quot;tool&quot; prompt, inject history<br>    history = &quot;\n&quot;.join(fashion_conversation.conversation_history)<br>    prompt = get_prompt(MODEL_TYPE, &quot;tool&quot;).format(history, &quot;User: &quot; + command)<br>    <br>    # Send command to the model<br>    output = llm(prompt, stop=[&quot;User:&quot;], max_tokens=800)<br>    response = output[&#39;choices&#39;][0][&#39;text&#39;]<br>    print (str(response))<br>    response = response.replace(&#39;\n&#39;, &#39;&#39;)<br>    # try to find json in the response string<br>    # heuristic needed mainly for Llama2 model responses<br>    # Llama2 usually responds like this: &quot;Sure! I can help in this, here&#39;s the json...&quot;<br>    try:<br>        # Extract json from model response by finding first and last brackets {}<br>        firstBracketIndex = response.index(&quot;{&quot;)<br>        lastBracketIndex = len(response) - response[::-1].index(&quot;}&quot;)<br>        jsonString = response[firstBracketIndex:lastBracketIndex]<br>        responseJson = json.loads(jsonString)<br>        marketplace_context = call_marketplace (responseJson)<br>        return marketplace_context, True<br>    except Exception as e:<br>        print(e)<br>    # No json match, just return False to denote that we should ignore this response<br>    return response, False<br><br><br>def agent_response(user_message):<br>    # process user message with the &quot;tool&quot; mode of LLM<br>    marketplace_context, blnNeedAPICall = check_tool(user_message)<br>        <br>    if blnNeedAPICall:<br>        history = &quot;\n&quot;.join(fashion_conversation.conversation_history)<br>        user_augmented_message = str(marketplace_context) + &quot;\n\n&quot; + &quot;User: &quot; + user_message<br>        current_context = str(marketplace_context)<br>        current_user_message = &quot;User: &quot; + user_message<br>        # Update &quot;chat&quot; prompt<br>        new_prompt = get_prompt(MODEL_TYPE, &quot;chat&quot;).format(history, current_user_message, current_context)<br>        fashion_conversation.add_message(&quot;User: &quot;, user_augmented_message)<br>    else:<br>        # if no function is selected to be called, totally ignore this response and re-prompt user message<br>        # to get the final response from the &quot;chat&quot; mode LLM<br>        current_user_message = &quot;User: &quot; + user_message<br>        history = &quot;\n&quot;.join(fashion_conversation.conversation_history)<br>        current_context = &quot;There is no context to provide. Answer User based on system instructions and current conversation history. &quot; #no context to provide, just use the history and respond<br>        new_prompt = get_prompt(MODEL_TYPE, &quot;chat&quot;).format(history, current_user_message, current_context)<br>        fashion_conversation.add_message(&quot;User: &quot;, user_message)<br>    <br>    # Send command to the model<br>    output = llm(new_prompt, stop=[&quot;User:&quot;], max_tokens=800)<br>    finalResponse = output[&#39;choices&#39;][0][&#39;text&#39;]<br>    #update the history<br>    fashion_conversation.add_message(&quot;Assistant: &quot;, finalResponse)<br>    return finalResponse</pre><p>The prompt for the “chat” mode is presented below. The history of the conversation is common for the 2 different modes of LLM usage. The Context is given after the user message, as if given before the model ignores it sometimes.</p><pre>chat_prompt = &#39;&#39;&#39;<br>      ### System:<br>      You are a helpful assistant of a women&#39;s fashion marketplace.<br>      Assistant recommends clothes to the user and gives information about the marketplace&#39;s designers and brands by ONLY using the Context provided.<br>      Assistant never responds with information that is not in the Context. <br>      Context is given in json format.<br>      Assistant is polite and always tries to sell products to the user.<br>      Assistant never shows the ID of products to the user.<br>      Assistant provides a textual description for each product that should only contain information provided in the json formatted Context about its Title, Description and Price and always be accompanied with its product url (product_url provided in Context).<br>      Assistant never invents new products not existing in the Context.<br>      Assistant only participates in conversations about fashion and kindly avoids to answer to every other request.<br><br>      The history of the current conversation between Assistant and User is:<br>      {0}<br>      <br>      ### User:<br>      {1}<br><br>      Context:<br>      {2}<br><br>      ### Assistant:<br>      <br>      &#39;&#39;&#39;</pre><p>So, did it work? Well, yes it did!!!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/723/1*P0TXa62NTg0q8p2_K2kpjA.png" /></figure><p>Intel’s new model performed surprisingly well. It proved to be extremely efficient. Very consistent in selecting the functions and returning structured output (function and parameters), highly capable of handling all difficult cases. Our 2-mode approach combined with the specific model can definitely be considered an alternative to Open AI Function Calling and models. It’s slightly more complicated as a solution, it may need a bit more tweaking and error handling, but it will work. There are so many examples on the internet using Langchain for agents, but actually few of them will work with open-source models, especially lighter ones in the category of 7 billion parameters. Being able to use a 7 billion parameter model with strong inference and reasoning capabilities as the brain of an agent can be a game changer.</p><p>Now, I’m really curious to also test this model in several ReAct scenarios…</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=242d771e7c79" width="1" height="1" alt=""><hr><p><a href="https://medium.com/11tensors/connect-an-ai-agent-with-your-api-intel-neural-chat-7b-llm-can-replace-open-ai-function-calling-242d771e7c79">Connect an AI agent with your API: Intel Neural-Chat 7b LLM can replace Open AI Function Calling</a> was originally published in <a href="https://medium.com/11tensors">11tensors</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Greek LLM after fine-tuning Llama2 7b]]></title>
            <link>https://medium.com/11tensors/a-greek-llm-after-fine-tuning-llama2-7b-678a8eca1ab3?source=rss-ffa5cd2a8118------2</link>
            <guid isPermaLink="false">https://medium.com/p/678a8eca1ab3</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llama-2]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Vassilios Antonopoulos]]></dc:creator>
            <pubDate>Wed, 06 Dec 2023 13:25:44 GMT</pubDate>
            <atom:updated>2023-12-13T13:43:59.000Z</atom:updated>
            <content:encoded><![CDATA[<p>As a language not widely spoken on an international scale, Greek often faces challenges in terms of digital representation and accessibility. In the era of LLMs, the existence of one tailored to the nuances of the Greek language, offering performance on par with models trained for more widely spoken languages, would be a resource that could facilitate the creation of diverse applications, ultimately fostering a more inclusive and dynamic digital landscape for the Greek-speaking population. So this post is the first of an upcoming series exploring available options and models towards the direction of building one or more specialized Greek LLMs.</p><p>Open AI’s models, both GPT-3.5 and GPT-4 do offer a LLM with advanced capabilities, even for the Greek language. But this comes with significant caveats. First of all, the cost associated with its usage is a highly prohibited factor for many applications and users. It’s a solution that cannot be widely adopted. And in many scenarios, lighter and more specialized LLMs (or even trained deep learning models) would do the job faster, cost-effectively and with better performance.</p><p><strong>11tensors</strong> possesses a very large corpus of around <strong>1.5 billion Greek documents</strong>, by far the largest available online Greek corpus. Building a model from scratch isn’t an easy task, it’s very resource intensive and requires a large investment. On the contrary, fine-tuning an existing foundation model with high quality task-specific training data is a process that we can follow and create specialized LLMs on the Greek language for various tasks.</p><p>Open source models are closing the gap with the closed ones. The collaborative efforts of the open-source community have led to the development of sophisticated models that rival their proprietary counterparts in terms of performance and versatility. Two models have gained the most traction lately: <strong>Llama2</strong> and <strong>Mistral</strong>.</p><p><strong>Mistral</strong> is a fantastic model, outperforming all models under 30B and even beating 70B models in certain benchmarks. But it doesn’t seem to have incorporated adequate amount of data in the Greek language during its training. The model seems incapable of successfully generating text in Greek and its instruct version doesn’t follow the instructions given to process Greek texts. So, unfortunately, we should leave this model out. At least its current version (upd: confirmed that Mistral’s new model Mixtral doesn’t support Greek either).</p><p>On the other hand, Meta’s Llama2 seem to be able to understand basic Greek language and concepts. This is something that can be observed with the large 70 billion parameter model. It will respond most of the times in English, it won’t be consistently correct, but it gives you the impression that it has understood your request expressed in Greek. As far as the smaller 7B model is concerned, it doesn’t seem to be able to generate Greek text and follow instructions. But this is the type of model we are mostly interested in. 7B category models are generally more useful and suitable for wider adoption. A model that when quantized can run on your desktop or laptop (even on CPU), a model performing extremely fast and highly efficiently is our primary focus. But it seems very difficult to get from Llama2–7b model a correct response or useful generated text in Greek.</p><p>Our feeling is that the model has captured during the pretraining procedure information that makes it capable of handling information in Greek language, but we need to further train and fine-tune the model to achieve the required performance.</p><p>In order to confirm our feeling, we have decided to start with the <strong>Llama2–7b-chat</strong> model and fine-tune it with Greek text on a specific task.</p><p>We created a dataset of 80k documents (public articles of Greek sites) properly transformed to train the foundation model on the following 2 tasks:</p><p>a) text generation: headline -&gt; story</p><p>b) summarization: story -&gt; headline</p><p>The model was trained with the Axolotl package, using Deepspeed for parallelization on a 4 x RTX4090 GPU instance on vast.ai for 12 hours for 1 epoch. QLoRA was used (all target modules were trained).</p><p>We can see some stats on the training and evaluation loss below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nSOljBJlE65hqhK5.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nA6lM65zPk5F5mq9.png" /></figure><p>The results of this fine-tuning process were surprisingly good! Unlike the Llama2–7b model, our model always now responds in Greek whenever it is asked in Greek. As it has not been trained with several instructions in Greek, system prompts and the typical prompting words (such as Question, Context, Response, Answer etc.) should be written in English for the model to respond well. Considering its new capabilities and initial performance evaluation, the current model seems to be a very good candidate for RAG (Retrieval Augmented Generation) and Summarization applications in the Greek language.</p><p>Some interesting sample responses are presented below. The results are really impressive considering the number of texts used for the fine tune, so the potential of a larger fine-tune, with several types of instructions, manually and synthetically extracted from our corpus seems very big. It will be done along with the creation of a benchmark as no one exists yet for the Greek language.</p><p>Sample 11tensors model responses:</p><p><em>Prompt: Read the text below and answer the question using only information from the text.</em></p><p><em>Κάθε φυσικό σύστημα περιέχει (ή εναλλακτικά αποθηκεύει) μία ποσότητα που ονομάζεται ενέργεια. Ενέργεια, συνεπώς, είναι η ικανότητα ενός σώματος ή συστήματος να παραγάγει έργο. Ανάλογα με τον τρόπο που έχει αποκτηθεί, ανταλλαχθεί ή αποθηκευτεί, μπορούμε να μιλήσουμε για πολλές μορφές ενέργειας: Μηχανική ενέργεια, που συνδυάζει την κινητική και τη δυναμική. Ηλεκτρομαγνητική ενέργεια, που συνδυάζει την ηλεκτρική και τη φωτεινή ή ενέργεια ακτινοβολίας, Πυρηνική ενέργεια Θερμική ενέργεια Χημική ενέργεια Υλο-ενέργεια.</em></p><p><em>Question: Πες μου 4 διαφορετικές μορφές ενέργειας.</em></p><p><em>Answer: Μηχανική ενέργεια, ηλεκτρομαγνητική ενέργεια, πυρηνική ενέργεια, θερμική ενέργεια</em></p><p><em>Prompt: Read the context below and answer the question using only the information provided in the context.</em></p><p><em>Context: Σύμφωνα με ανακοίνωση των αρχών άγνωστοι δράστες, που ήταν ένοπλοι, επιτέθηκαν σε χρηματαποστολή επιχείρησης, η οποία ανήκει στην οικογένεια της συζύγου του Λιονέλ Μέσι, Αντονέλα. Οι δράστες επιτέθηκαν στο αυτοκίνητο που μετέφερε δύο υπαλλήλους του σούπερ μάρκετ, καθώς και τη ξαδέρφη της αγαπημένης του Αργεντινού σταρ, σπάζοντας τα παράθυρα. Αφού το διέρρηξαν, αφαίρεσαν δύο τσάντες που περιείχαν εισπράξεις ύψους 19.700 ευρώ, τα οποία επρόκειτο να κατατεθούν σε τράπεζα σε κεντρικό σημείο του Ροζάριο.</em></p><p><em>Question: Πόσα χρήματα είχαν οι τσάντες που πήραν οι διαρρήκτες? Answer: 19.700 ευρώ</em></p><p><em>Prompt: Γράψε έναν τίτλο για την ακόλουθη ιστορία:</em></p><p><em>Με 37 έδρες σε σύνολο 150 της βουλής, το ισλαμοφοβικό ακροδεξιό κόμμα του Γκέερτ Βίλντερς αναδείχθηκε πρώτη δύναμη στις χθεσινές εκλογές στην Ολλανδία, προκαλώντας έναν σεισμό που κινδυνεύει να γίνει αισθητός σε όλη την Ευρώπη. Δεν είναι βέβαια ακόμα σαφές εάν ο Βίλντερς θα μπορέσει πραγματικά να σχηματίσει κυβέρνηση, καθώς δεν αποκλείεται να συγκροτηθεί ένας ευρύς συνασπισμός των υπόλοιπων κομμάτων εναντίον του.</em></p><p><em>Response: Ολλανδία: Αναδείχθηκε πρώτη δύναμη το ισλαμοφοβικό ακροδεξιό κόμμα του Βίλντερς</em></p><p><em>Prompt: Answer in Greek the following question.</em></p><p><em>Question: Ποιες είναι οι 5 μεγαλύτερες ελληνικές πόλεις;</em></p><p><em>Answer: Αυτές είναι οι 5 μεγαλύτερες ελληνικές πόλεις: 1. Αθήνα 2. Θεσσαλονίκη 3. Πάτρα 4. Λάρισα 5. Κοζάνη</em></p><p><em>Prompt: Read the text below and answer the question using only information from the text.</em></p><p><em>Context: Ανακοινώθηκε οι διαιτητές της 12ης αγωνιστικής. Στο ματς της ΑΕΚ με τον ΠΑΣ στα Γιάννενα ορίστηκε ο Διαμαντόπουλος με VAR τον Μανούχος. Ο Πορτογάλος Γκουστάβο Κορέια θα διευθύνει το Άρης-Παναθηναϊκός. Πιο αναλυτικά: 16.00 Κηφισιά — Αστέρας Τρίπολης: Διαιτητής: Τσακαλίδης, Βοηθοί: Παπαδάκης, Κομισοπούλου , 4ος: Βάτσιος , VAR: Παπαπέτρου , AVAR: Φωτόπουλος</em></p><p><em>17.00 ΠΑΣ Γιάννινα — ΑΕΚ: Διαιτητής: Διαμαντόπουλος, Βοηθοί: Ψάρρης, Τριανταφυλλου, 4ος: Μαλούτας, VAR: Μανούχος , AVAR: Μεϊντανάς.</em></p><p><em>19.30 Λαμία — Ατρόμητος: Διαιτητής: Βεργέτης, Βοηθοί: Στεφανής, Κορώνας, 4ος: Φωτιάς, VAR: Τσαγκαράκης, AVAR:Σαμοϊλης.</em></p><p><em>20.30 Άρης — Παναθηναϊκός: Διαιτητής: Κορέια , Βοηθοί: Τιάγκο, Tεϊξέιρα, 4ος: Τσιάρας, VAR: Εστέβες , AVAR: Nικολακάκης.</em></p><p><em>Question: Ποιος διαιτητής ορίστηκε στον αγώνα της ΑΕΚ?</em></p><p><em>Answer: Διαιτητής ορίστηκε στον αγώνα της ΑΕΚ ο Διαμαντόπουλος</em></p><p><em>Prompt: Answer the question below.</em></p><p><em>Question: Πόσα γραμμάρια έχει το 1 κιλό?</em></p><p><em>Answer: 1000</em></p><p><em>(1-shot example needed to succeed here)</em></p><p><em>Prompt: Answer the following question.</em></p><p><em>Question: Ποια είναι η πρωτεύουσα του νομού Αρκαδίας; Answer: Η η πρωτεύουσα του νομού Αρκαδίας είναι η Τρίπολη.</em></p><p><em>Question: Ποια είναι η πρωτεύουσα του νομού Μεσσηνίας; Answer: Η πρωτεύουσα του νομού Μεσσηνίας είναι η Καλαμάτα.</em></p><p><em>Prompt: Read the context below and answer the question using only the information provided in the context.</em></p><p><em>Context: Στο πλαίσιο του αυτοφώρου, συνελήφθησαν την Τρίτη στον Κολωνό τέσσερις ανήλικοι, ηλικίας από 13 έως 15 ετών, καθώς κατηγορούνται- κατά περίπτωση- για εξύβριση, απειλή και απλές σωματικές βλάβες. Ειδικότερα, όπως έγινε γνωστό από την Αστυνομία χθες το μεσημέρι στο πάρκο της Ακαδημίας Πλάτωνος και ενώ ήταν σε εξέλιξη εκδρομή μαθητών σχολείων δευτεροβάθμιας εκπαίδευσης διαφορετικών περιοχών, οι κατηγορούμενοι επιτέθηκαν με υβριστικές εκφράσεις, απειλές και χρήση σωματικής βίας σε βάρος πέντε μαθητών (τέσσερα αγόρια και κορίτσι ηλικίας από 16 έως 17 ετών) του άλλου σχολείου. Στο σημείο μετέβη ασθενοφόρο, το πλήρωμα του οποίου παρείχε τις πρώτες βοήθειες στους τραυματισμένους μαθητές και οι αστυνομικοί της Άμεσης Δράσης που κλήθηκαν και έσπευσαν στο σημείο, συνέλαβαν καθ’ υπόδειξη των αρμόδιων καθηγητών τους μαθητές που βιαιοπράγησαν.</em></p><p><em>Question: Πόσoι ανήλικοι συνελήφθησαν στον Κολωνό?</em></p><p><em>Answer: Συνελήφθησαν τέσσερις ανήλικοι</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=678a8eca1ab3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/11tensors/a-greek-llm-after-fine-tuning-llama2-7b-678a8eca1ab3">A Greek LLM after fine-tuning Llama2 7b</a> was originally published in <a href="https://medium.com/11tensors">11tensors</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Migrating our chatbot from Beep Boop to Heroku]]></title>
            <link>https://medium.com/qualia-ai/migrating-our-chatbot-from-beep-boop-to-heroku-698c6c7b7559?source=rss-ffa5cd2a8118------2</link>
            <guid isPermaLink="false">https://medium.com/p/698c6c7b7559</guid>
            <category><![CDATA[beep-boop]]></category>
            <category><![CDATA[heroku]]></category>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[slack]]></category>
            <dc:creator><![CDATA[Vassilios Antonopoulos]]></dc:creator>
            <pubDate>Wed, 22 Nov 2017 10:22:40 GMT</pubDate>
            <atom:updated>2017-11-22T10:22:40.924Z</atom:updated>
            <content:encoded><![CDATA[<p>We loved Beep Boop from the very beginning. It came as a natural fit for us at just the right time in order to ease and accelerate our Slack development. Although we were very early users of Beep Boop, we felt really confident to move on with them, once we understood how devoted and capable the team was. Their platform quickly became very stable and so Brad very soon stopped getting my messages first thing in the morning before his coffee (due to our time difference, I was sometimes the first one to alert them for possible issues 😊). Moreover, they had always been willing to help us with any development issues and decisions we needed to take for our own project. And that was really precious!</p><p>Quite recently Beep Boop <a href="https://blog.beepboophq.com/the-final-chapter-of-beep-boop-efab4f351a51">announced </a>that it is coming to an end. Hmm… bad news 😞 . We are definitely unhappy with this, but once they decided that <a href="https://missions.ai/"><strong>Missions</strong></a> is the way to go… Go for it guys 🎯! There is some good news though. They will keep supporting and developing the Slapp module for Slack. Botkit proved to be a headache and once we switched to <strong>Slapp </strong>and took advantage of the Slack Events API everything just started working like a charm.</p><p>So, where should we now host <strong>qualiabot</strong> (this is the name of our Slack bot)? Well, we were already familiar with <strong>Heroku</strong>, so we decided to switch to it. The overall procedure of migrating from Beep Boop to Heroku has been straightforward. The sample code provided by Brad was very helpful, but once I was asked to write about it, by some members of the Beep Boop community, I decided to create this post as a time saver for them. Below are the exact steps of this procedure in some detail.</p><p>A short introduction about our chatbot: It is implemented in Node.js using the Slapp module (so it uses the Slack Events API) and an Express web server. It’s a multi –team bot located in the Slack App Directory. Our clients install it to their teams and gain access to online monitoring and analytics insights, which they can request in natural language without using our web UI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/470/1*SwZkkMLiXjJAeF8tHo4sDg.png" /></figure><p>So, briefly here is what you will have to do:</p><ol><li>Your bot may or may not already be using a <strong>database </strong>for storing your application’s metadata. Well, in any case, now you will need one database for sure, in order to handle the stuff that Beep Boop silently took care for you until now.</li><li>Some Teams have already installed your bot and Beep Boop manages their metadata so far. You should now <strong>migrate </strong>these metadata into your own database.</li><li>Slapp needs access to Context (Team) and Conversational (state) data in order to apply its magic. This has been also taken care of Beep Boop. Now you have to do it. But don’t panic, it’s easy. They have published a sample project in github that shows a few functions your <strong>database module should implement</strong> in order to achieve this.</li><li>For the new teams that will install your bot in the future: Replace BeepBoop<strong> ‘Add to Slack’</strong> button with the original one offered by Slack and handle the OAUTH flow by yourself, saving the data returned by Slack during this process into your database.</li><li>Go to your Slack app page and <strong>change the Event API url</strong>, so that it now points to your new location. This way, you let Slack know that the Events should be sent to your new location and not to Beep Boop.</li></ol><blockquote><strong>The Heroku part</strong></blockquote><p>Before I walk you through this process in more detail, let’s briefly surface the <strong>Heroku </strong>part. You will need at least one dyno. As free dynos sleep after a short inactivity period and do not provide (free) SSL, I would propose to start with a Hobby dyno for the tests and if more processing power is needed, you can upgrade or scale accordingly. Download Heroku CLI and setup a git remote for your project in order to be able to upload the code. Add SSL support to your application in the Settings page and also configure there all your environment variables. Heroku rebuilds your project. Add the version of the node in the <strong>package.json</strong> file of your project, under the <strong>“engines”</strong> element (e.g. “node” : ”6.12.0”), so that it knows you are uploading a node.js project. Also be careful with the dependencies in this file, you may need to explicitly define their versions, as Heroku will download the latest available version and of course this may break your code. Moreover, you need to add the <strong>Procfile</strong>. One line of code, such as “web: node server.js”, where <em>server.js</em> is your starting point and it is fine. It is critical to be aware of the fact that in Heroku you <strong>don’t have access to the file system</strong>. If you were storing data in the file system in Beep Boop, you will have to change that part of your code and use a database instead.</p><p>Let’s see now the migration procedure in more detail. Throughout of the text I will be referring to the <strong>slapp-firebase-example</strong>, which you can find <a href="https://github.com/BeepBoopHQ/slapp-firebase-example">here</a>.</p><p><a href="https://github.com/BeepBoopHQ/slapp-firebase-example">BeepBoopHQ/slapp-firebase-example</a></p><blockquote><strong>Decide on the database to use</strong></blockquote><p>If you already have a database, you may of course extend its usage and also store there the required Team &amp; Conversational metadata. If you don’t have one, select one you are familiar with. All you will have to do is to implement a few basic functions that Slapp expects from your database module. You may go for Google’s Firebase database, for which the implementation is ready in the <strong>slapp-firebase-example</strong> module.</p><blockquote><strong>Migrate Team Data</strong></blockquote><p>First of all, retrieve all the current data of the teams that have installed your Beep Boop bot.</p><p>You will need the BeepBoop API Token. It exists on the Settings tab of your bot’s page in BeepBoop.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/557/1*VDx9USCH4OoKAUoNyDAJBA.png" /></figure><p>Then use it as the authorization parameter of a GET request to <a href="https://beepboophq.com/api/v1/slack-teams">https://beepboophq.com/api/v1/slack-teams</a> (see below).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/499/1*pth46HKpzzXrBGyE-C563Q.png" /></figure><p>The json result contains all the data that you need. Now you have to transform this data to a format that is compatible with Slack’s OAUTH response. I have prepared a script in Node.js that will transform the previous json using the <strong><em>node-json-transform</em></strong> module:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b8427227fbfa9db85bd0002cc071141c/href">https://medium.com/media/b8427227fbfa9db85bd0002cc071141c/href</a></iframe><p>The output is a json file, which is ready to be imported into your database. You should insert each value as a separate record/document in your database in a table/collection named “<strong>teams”</strong>.</p><blockquote><strong>Connect Slapp with your database</strong></blockquote><p>A few lines of code are all that it is needed in order to connect to the database and make some simple statements. An example of what is required for the Google’s Firebase database is contained in the <em>db.js</em> module of the slapp-firebase-example<em>.</em> One should implement the functions saveTeam, getTeam, saveConvo, getConvo and deleteConvo, which are required from Slapp, as shown in that example. Then pass the db object as a parameter to your Slapp object during its creation.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/961eb396b3d80e968466998d28042434/href">https://medium.com/media/961eb396b3d80e968466998d28042434/href</a></iframe><blockquote><strong>Replace ‘Add to Slack’ button</strong></blockquote><p>Time to get ready for your new users. Replace BeepBoop ‘Add to Slack’ Button with the one that Slack contains in your app’s page in <a href="https://api.slack.com/apps">https://api.slack.com/apps</a> (Manage Distribution menu). Depending on your OAUTH flow of course, you may need to add the <strong>redirect_uri</strong> parameter along with <strong>scope</strong> and <strong>client_id</strong>.</p><p>Then, on your web server, you should handle the GET request to the mydomain/Add-to-Slack. Test the token and save the team data in your database (remember the .saveTeam() function you implemented earlier in db module). An explanatory example is contained in the oauth.js file of the <strong>slapp-firebase-example</strong> module that shows exactly how it is done.</p><blockquote><strong>OK, Slack. Stop sending the Events to Beep Boop!</strong></blockquote><p>Slack should now stop sending the Events to the BeepBoop url you have defined in the past. Go to <a href="https://api.slack.com/apps/">https://api.slack.com/apps/</a>, first select your app and then the Events Subscription menu item. You should now change it to something like <a href="https://mydomain.herokuapp.com/slack/event">https://mydomain.herokuapp.com/slack/event</a> (or if you have set a custom domain <a href="https://mycustomdomain/slack/event">https://mycustomdomain/slack/event</a>). Of course you will need SSL on the server, but even Hobby dynos offer it for free now in Heroku. Don’t be in a rush! If you just tried it, you may have seen the red failure message J ! In order for the change to succeed, your web server should respond to a specific Slack event with a challenge value J… All you need to do is implement the following handle for this Event using Slapp.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/638182fbe27ed09a355dde833da7d3a7/href">https://medium.com/media/638182fbe27ed09a355dde833da7d3a7/href</a></iframe><p>Now, if you try to change the url, you should succeed. Press the <strong>SAVE</strong> button on the Slack page in order to store this new url. You should now start getting Slack Events into your Heroku instance.</p><p>But be careful! Until now, you have a fully working development app. If your app resides in the Slack App Directory, in order for this Events API url change to take effect in your production instance, you will need to submit your app once again for Review.</p><blockquote><strong>Submit the Slack app for review</strong></blockquote><p>Don’t worry. Preparing this submission is fast. All you need is click on all required boxes (nothing has changed from the previous submission in these general terms). In the Notes textbox, a message like <em>“Hey folks, I have changed the Events API url from BeepBoop to Heroku and I need my production instance to be updated.”</em> should be more than enough. Within few hours you will get the approval.</p><p>As a conclusion, Brad’s github example is very helpful and is the main resource you will have to visit while migrating from BeepBoop. But I also hope that this post will save you some time making the steps of the whole procedure more clear and emphasizing on some details here and there. Your bot will stay alive 💪 💪 💪!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=698c6c7b7559" width="1" height="1" alt=""><hr><p><a href="https://medium.com/qualia-ai/migrating-our-chatbot-from-beep-boop-to-heroku-698c6c7b7559">Migrating our chatbot from Beep Boop to Heroku</a> was originally published in <a href="https://medium.com/qualia-ai">Qualia.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>