first commit

.gitignore

@@ -0,0 +1,8 @@
+.env
+__pycache__/app.cpython-38.pyc
+__pycache__/app.cpython-39.pyc
+__pycache__/utils.cpython-38.pyc
+
+notebooks/
+*.pyc
+local_tests/

README.md

@@ -7,6 +7,9 @@ sdk: gradio
 sdk_version: 4.36.1
 app_file: app.py
 pinned: false
+hf_oauth: true
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Spinoza Project
+
+WIP

app.py

@@ -0,0 +1,582 @@
+import time
+import yaml
+import logging
+import gradio as gr
+from langchain.prompts.chat import ChatPromptTemplate
+from huggingface_hub import hf_hub_download, whoami
+from app.source.backend.llm_utils import get_llm
+from app.source.backend.document_store import pickle_to_document_store
+from app.source.backend.get_prompts import get_qa_prompts
+from app.source.frontend.utils import (
+    make_html_source,
+    make_html_presse_source,
+    init_env,
+)
+from app.source.backend.prompt_utils import to_chat_instruction, SpecialTokens
+
+init_env()
+
+with open("./app/config.yaml") as f:
+    config = yaml.full_load(f)
+
+prompts = {}
+for source in config["prompt_naming"]:
+    with open(f"./app/prompt_{source}.yaml") as f:
+        prompts[source] = yaml.full_load(f)
+
+## Building LLM
+print("Building LLM")
+model = "gpt35turbo"
+llm = get_llm()
+
+## Loading_tools
+print("Loading Databases")
+qdrants = {
+    tab: pickle_to_document_store(
+        hf_hub_download(
+            repo_id="momenaca/spinoza-dbs",
+            filename=f"database_{tab}.pickle",
+            token=True,
+            repo_type="dataset",
+        )
+    )
+    for tab in config["prompt_naming"]
+}
+
+## Load Prompts
+print("Loading Prompts")
+chat_qa_prompts, chat_reformulation_prompts, chat_summarize_memory_prompts = {}, {}, {}
+for source, prompt in prompts.items():
+    chat_qa_prompt, chat_reformulation_prompt = get_qa_prompts(config, prompt)
+    chat_qa_prompts[source] = chat_qa_prompt
+    chat_reformulation_prompts[source] = chat_reformulation_prompt
+    # chat_summarize_memory_prompts[source] = chat_summarize_memory_prompt
+
+with open("./assets/style.css", "r") as f:
+    css = f.read()
+
+
+def update_tabs(outil, visible_tabs):
+    visible_tabs = outil
+    return visible_tabs
+
+
+special_tokens = SpecialTokens(config)
+
+synthesis_template = """You are a factual journalist that summarize the secialized awnsers from thechnical sources.
+
+Based on the folowing question:
+{question}
+
+And the following expert answer:
+{answers}
+
+Answer the question, in French.
+When using legal awnsers, keep tracking of the name of the articles.
+When using ADEME awnsers, name the sources that are mainly used.
+List the different element mentionned, and highlight the agreement points between the sources, as well as the contradictions or differences.
+Generate the answer as markdown, with an aerated layout, and headlines in bold
+Start by a general summary, agreement and contracdiction, and then go into detail without paraphasing the experts awnsers.
+"""
+
+synthesis_prompt = to_chat_instruction(synthesis_template, special_tokens)
+synthesis_prompt_template = ChatPromptTemplate.from_messages([synthesis_prompt])
+
+
+def zip_longest_fill(*args, fillvalue=None):
+    # zip_longest('ABCD', 'xy', fillvalue='-') --> Ax By C- D-
+    iterators = [iter(it) for it in args]
+    num_active = len(iterators)
+    if not num_active:
+        return
+
+    cond = True
+    fillvalues = [None] * len(iterators)
+    while cond:
+        values = []
+        for i, it in enumerate(iterators):
+            try:
+                value = next(it)
+            except StopIteration:
+                value = fillvalues[i]
+            values.append(value)
+
+        new_cond = False
+        for i, elt in enumerate(values):
+            if elt != fillvalues[i]:
+                new_cond = True
+        cond = new_cond
+
+        fillvalues = values.copy()
+        yield tuple(values)
+
+
+def build_data_dict(config):
+    data_dict = {}
+    for tab in config["tabs"]:
+        data_dict[tab] = {
+            "tab": {
+                "init_value": tab,
+                "component": None,
+                "elem_id": "tab",
+            },
+            "description": {
+                "init_value": config["tabs"][tab],
+                "component": None,
+                "elem_id": "desc",
+            },
+            "question": {
+                "init_value": None,
+                "component": None,
+                "elem_id": "question",
+            },
+            "answer": {
+                "init_value": None,
+                "component": None,
+                "elem_id": "answer",
+            },
+            "sources": {
+                "init_value": None,
+                "component": None,
+                "elem_id": "src",
+            },
+        }
+    return data_dict
+
+
+def init_gradio(data, config=config):
+    for t in data:
+        data[t]["tab"]["component"] = gr.Tab(
+            data[t]["tab"]["init_value"], elem_id="tab"
+        )
+        with data[t]["tab"]["component"]:
+            for fields in data[t]:
+                if fields == "question":
+                    data[t][fields]["component"] = gr.Textbox(
+                        elem_id=data[t][fields]["elem_id"],
+                        show_label=False,
+                        interactive=True,
+                        placeholder="",
+                    )
+                # elif fields == "answer":
+                #     data[t][fields]["component"] = gr.Textbox(
+                #         elem_id=data[t][fields]["elem_id"],
+                #         show_label=True,
+                #         interactive=True,
+                #         placeholder="",
+                #         show_copy_button=True
+                #     )
+                elif fields != "tab":
+                    data[t][fields]["component"] = gr.Markdown(
+                        data[t][fields]["init_value"],
+                        elem_id=data[t][fields]["elem_id"],
+                    )
+                    # data[t][fields]["component"] = gr.Textbox(
+                    #     value=data[t][fields]["init_value"],
+                    #     elem_id=data[t][fields]["elem_id"],
+                    #     show_label=True,
+                    #     interactive=False,
+                    #     show_copy_button=True,
+                    # )
+    return data
+
+
+def add_warning():
+    return "*Les éléments cochés ont commencé à être généré dans les onglets spécifiques, la synthèse ne sera disponible qu'après la mise à disposition de ces derniers.*"
+
+
+def format_question(question):
+    return f"{question}"  # ###
+
+
+def parse_question(question):
+    x = question.replace("<p>", "").replace("</p>\n", "")
+    if "### " in x:
+        return x.split("### ")[1]
+    return x
+
+
+def reformulate(outils, question, tab, config=config):
+    if tab in outils:
+        return llm.stream(
+            chat_reformulation_prompts[config["source_mapping"][tab]],
+            {"question": parse_question(question)},
+        )
+    else:
+        return iter([None] * 5)
+
+
+def reformulate_single_question(outils, question, tab, config=config):
+    for elt in reformulate(outils, question, tab, config=config):
+        time.sleep(0.02)
+        yield elt
+
+
+def reformulate_questions(outils, question, config=config):
+    for elt in zip_longest_fill(
+        *[reformulate(outils, question, tab, config=config) for tab in config["tabs"]]
+    ):
+        time.sleep(0.02)
+        yield elt
+
+
+def add_question(question):
+    return question
+
+
+def answer(question, source, outils, tab, config=config):
+    if tab in outils:
+        if len(source) < 10:
+            return iter(["Aucune source trouvée, veuillez reformuler votre question"])
+        else:
+
+            return llm.stream(
+                chat_qa_prompts[config["source_mapping"][tab]],
+                {
+                    "question": parse_question(question),
+                    "sources": source.replace("<p>", "").replace("</p>\n", ""),
+                },
+            )
+    else:
+        return iter([None] * 5)
+
+
+def answer_single_question(outils, source, question, tab, config=config):
+    for elt in answer(question, source, outils, tab, config=config):
+        time.sleep(0.02)
+        yield elt
+
+
+def answer_questions(outils, *questions_sources, config=config):
+
+    questions = [elt for elt in questions_sources[: len(questions_sources) // 2]]
+    sources = [elt for elt in questions_sources[len(questions_sources) // 2 :]]
+
+    for elt in zip_longest_fill(
+        *[
+            answer(question, source, outils, tab, config=config)
+            for question, source, tab in zip(questions, sources, config["tabs"])
+        ]
+    ):
+        time.sleep(0.02)
+        yield elt
+
+
+def get_source_link(metadata):
+    return metadata["file_url"] + f"#page={metadata['content_page_number'] + 1}"
+
+
+def get_button(i, tag):
+    return f"""<button id="btn_{tag}_{i}" type="button" style="margin: 0; display: inline; align="right">[{i}]</button>"""
+
+
+def get_html_sources(buttons, cards):
+    return f"""
+<p style="margin: 0; display: inline;"><strong><br>Sources utilisées : </strong></p>
+{buttons}
+{cards}
+"""
+
+
+def get_sources(outils, question, tab, qdrants=qdrants, config=config):
+    k = config["num_document_retrieved"]
+    min_similarity = config["min_similarity"]
+    if tab in outils:
+        sources = qdrants[
+            config["source_mapping"][tab]
+        ].similarity_search_with_relevance_scores(
+            config["query_preprompt"]
+            + question.replace("<p>", "").replace("</p>\n", ""),
+            k=k,
+            # filter=get_qdrant_filters(filters),
+        )
+        sources = [(doc, score) for doc, score in sources if score >= min_similarity]
+
+        buttons_ids = list(range(len(sources)))
+        buttons = " ".join(
+            [get_button(i, tab) for i, source in zip(buttons_ids, sources)]
+        )
+        formated = (
+            "\n\n".join(
+                [
+                    make_html_presse_source(source[0], i, tab, source[1], config)
+                    for i, source in zip(buttons_ids, sources)
+                ]
+            )
+            if tab == "Presse"
+            else "\n\n".join(
+                [
+                    make_html_source(source[0], i, tab, source[1], config)
+                    for i, source in zip(buttons_ids, sources)
+                ]
+            )
+        )
+        formated = get_html_sources(buttons, formated) if sources else ""
+        text = "\n\n".join(
+            [
+                f"Doc {str(i)} with source type {elt[0].metadata.get('file_source_type')}:\n"
+                + elt[0].page_content
+                for i, elt in enumerate(sources)
+            ]
+        )
+        return str(formated), str(text)  # formated_sources, text_sources
+    else:
+        return "", ""
+
+
+def retrieve_sources(outils, *questions, qdrants=qdrants, config=config):
+    results = [
+        get_sources(outils, question, tab, qdrants, config)
+        for question, tab in zip(questions, config["tabs"])
+    ]
+    formated_sources = [source[0] for source in results]
+    text_sources = [source[1] for source in results]
+    return tuple(formated_sources + text_sources)
+
+
+def get_experts(outils, *answers, config=config):
+    return "\n\n".join(
+        [
+            f"{tab}\n{answers[i]}"
+            for i, tab in enumerate(config["tabs"])
+            if (tab in outils)
+        ]
+    )
+
+
+def get_synthesis(outils, question, *answers, config=config):
+    answer = []
+    for i, tab in enumerate(config["tabs"]):
+        if (tab in outils) & (len(str(answers[i])) >= 100):
+            answer.append(
+                f"{tab}\n{answers[i]}".replace("<p>", "").replace("</p>\n", "")
+            )
+
+    if len(answer) == 0:
+        return "Aucune source n'a pu être identifiée pour répondre, veuillez modifier votre question"
+    else:
+        for elt in llm.stream(
+            synthesis_prompt_template,
+            {
+                "question": question.replace("<p>", "").replace("</p>\n", ""),
+                "answers": "\n\n".join(answer),
+            },
+        ):
+            time.sleep(0.01)
+            yield elt
+
+
+def get_listener():
+    return """
+    function my_func_body() {
+        const body = document.querySelector("body");
+        body.addEventListener("click", e => {
+            console.log(e)
+            const sourceId = "btn_" + e.target.id.split("_")[1] + "_" + e.target.id.split("_")[2] + "_source"
+            console.log(sourceId)
+                if (document.getElementById(sourceId).style.display === "none") {
+                document.getElementById(sourceId).style.display = "";
+            } else {
+                document.getElementById(sourceId).style.display = "none";
+            }
+        }
+    )}
+    """
+
+
+def get_source_template(buttons, divs_source):
+    return """
+    <div class="source">
+        <p style="margin: 0; display: inline;"><strong><br>Sources utilisées :</strong></p>
+        {buttons}
+        {divs_source}
+        </div>
+    </div>
+    """
+
+
+def activate_questions(outils, *textboxes, config=config):
+    activated_textboxes = []
+    for i, tab in enumerate(config["tabs"]):
+        if tab in outils:
+            activated_textboxes.append(
+                gr.Textbox(
+                    show_label=False,
+                    interactive=True,
+                    placeholder="Sélectionnez cet outil et posez une question sur l'onglet de synthèse",
+                )
+            )
+
+        else:
+            activated_textboxes.append(
+                gr.Textbox(
+                    show_label=False,
+                    interactive=False,
+                    placeholder="Sélectionnez cet outil et posez une question sur l'onglet de synthèse",
+                )
+            )
+    return activated_textboxes
+
+
+def empty():
+    return ""
+
+
+def empty_none():
+    return None
+
+
+theme = gr.themes.Soft(
+    primary_hue="sky",
+    font=[gr.themes.GoogleFont("Poppins"), "ui-sans-serif", "system-ui", "sans-serif"],
+)
+
+logo_rsf = config["logo_rsf"]
+logo_ap = config["logo_ap"]
+
+data = build_data_dict(config)
+
+
+def update_visible(oauth_token: gr.OAuthToken | None):
+    if oauth_token is None:
+        return {
+            bloc_1: gr.update(visible=True),
+            bloc_2: gr.update(visible=False),
+            bloc_3: gr.update(visible=False),
+        }
+
+    org_names = [org["name"] for org in whoami(oauth_token.token)["orgs"]]
+
+    if "SpinozaProject" in org_names:  # logged in group
+        return {
+            bloc_1: gr.update(visible=False),
+            bloc_2: gr.update(visible=True),
+            bloc_3: gr.update(visible=False),
+        }
+
+    else:  # logged but not in group
+        return {
+            bloc_1: gr.update(visible=False),
+            bloc_2: gr.update(visible=False),
+            bloc_3: gr.update(visible=True),
+        }
+
+
+with gr.Blocks(
+    title=f"🔍{config['demo_name']}",
+    css=css,
+    js=get_listener(),
+    theme=theme,
+) as demo:
+    gr.LoginButton()
+
+    with gr.Column() as bloc_1:
+        textbox_1 = gr.Textbox("You are not logged to Hugging Face !", show_label=False)
+
+    with gr.Column(visible=False) as bloc_3:
+        textbox_3 = gr.Textbox(
+            "You are not part of the Spinoza Project, ask access here : https://huggingface.co/organizations/TestSpinoza/share/kmwhyFXasNnGfkBrKzNAPgnlRrxyVOSSMx"
+        )
+
+    with gr.Column(visible=False) as bloc_2:
+        gr.HTML(
+            f"""<div class="row_logo">
+                        <img src={logo_rsf} alt="logo RSF" style="float:left; width:120px; height:70px">
+                        <img src={logo_ap} alt="logo AP" style="width:120px; height:70px">
+                </div>"""
+        )
+
+        text_sources = {elt: gr.State("") for elt in config["tabs"]}
+        tab_states = {elt: gr.State(elt) for elt in config["tabs"]}
+        with gr.Row():
+            with gr.Column(scale=3):
+                outils = gr.CheckboxGroup(
+                    choices=list(config["tabs"].keys()),
+                    value=list(config["tabs"].keys()),
+                    type="value",
+                    label="Choisir les bases de données à interroger",
+                )
+            with gr.Column(scale=1):
+                submit_btn = gr.Button(
+                    "Relancer la Synthèse", variant="primary", elem_id="synthese_btn"
+                )
+
+        # Synthesis tab
+        synthesis_tab = gr.Tab("Synthesis", elem_id="tab")
+        with synthesis_tab:
+            question = gr.Textbox(
+                show_label=True,
+                label="Posez une question à Spinoza",
+                placeholder="Quelle est votre question ?",
+            )
+            md_question = gr.Markdown(None, visible=False)
+            warning = gr.Markdown(None, elem_id="warn")
+            synthesis = gr.Markdown(None, elem_id="synthesis")
+
+        data = init_gradio(data)
+        (
+            question.submit(add_question, [question], [md_question])
+            .then(add_warning, [], [warning])
+            .then(empty, [], [synthesis])
+            .then(
+                reformulate_questions,
+                [outils, md_question],
+                [data[tab]["question"]["component"] for tab in config["tabs"]],
+            )
+            .then(
+                retrieve_sources,
+                [outils]
+                + [data[tab]["question"]["component"] for tab in config["tabs"]],
+                [data[tab]["sources"]["component"] for tab in config["tabs"]]
+                + [text_sources[tab] for tab in config["tabs"]],
+            )
+            .then(
+                answer_questions,
+                [outils]
+                + [data[tab]["question"]["component"] for tab in config["tabs"]]
+                + [text_sources[tab] for tab in config["tabs"]],
+                [data[tab]["answer"]["component"] for tab in config["tabs"]],
+            )
+            .then(
+                get_synthesis,
+                [outils, md_question]
+                + [data[tab]["answer"]["component"] for tab in config["tabs"]],
+                [synthesis],
+            )
+        )
+
+        for tab in config["tabs"]:
+            (
+                data[tab]["question"]["component"]
+                .submit(empty, [], [data[tab]["sources"]["component"]])
+                .then(empty, [], [text_sources[tab]])
+                .then(empty, [], [data[tab]["answer"]["component"]])
+                .then(
+                    get_sources,
+                    [outils, data[tab]["question"]["component"], tab_states[tab]],
+                    [data[tab]["sources"]["component"], text_sources[tab]],
+                )
+                .then(
+                    answer_single_question,
+                    [
+                        outils,
+                        text_sources[tab],
+                        data[tab]["question"]["component"],
+                        tab_states[tab],
+                    ],
+                    [data[tab]["answer"]["component"]],
+                )
+            )
+
+        (
+            submit_btn.click(empty, [], [synthesis]).then(
+                get_synthesis,
+                [outils, md_question]
+                + [data[tab]["answer"]["component"] for tab in config["tabs"]],
+                [synthesis],
+            )
+        )
+    demo.load(update_visible, inputs=None, outputs=[bloc_1, bloc_2, bloc_3])
+
+if __name__ == "__main__":
+    demo.queue().launch(share=True, debug=True)

app/config.yaml

@@ -0,0 +1,46 @@
+demo_name: Spinoza Q&A
+tabs:
+  GIEC et IPBES: "*Outil dédié aux rapports du GIEC et de l'IPBES.*"
+  Textes Juridiques: "*Outil dédié aux codes Français modifiés par la loi climat (21/73).*"
+  Documents Stratégiques: "*Outil dédié aux données centrées sur le plan politique (SNBC).*"
+  ADEME: "*Outil dédié aux données issues de l'ADEME et nous avons sélectionnés notamment différentes catégories de rapports:*\n
+              * *Les guides mis à disposition de la population*\n
+              * *Les rapport d'expériences sur des nouvelles technologies*\n
+              * *Des études et recherches sur des impacts locaux*\n
+              * *Des documents institutionnels (analyses demandées par la France & rapports d'activité)*\n
+              * *Les plans de transition sectoriels pour les secteurs industriels les plus émetteurs : (verre, papier, ciment, acier, aluminium, chimie, sucre)*"
+  # Presse: "*Outil dédié aux données fournies par Aday concernant la presse.*"
+
+logo_rsf: 
+  "data:image/png;base64,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"
+ 
+logo_ap:
+   "data:image/png;base64,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"
+
+source_mapping:
+  GIEC et IPBES: "Science"
+  Textes Juridiques: "Loi"
+  Documents Stratégiques: "Politique"
+  ADEME: "ADEME"
+  # Presse: "Presse"
+
+prompt_naming:
+  Science: "Science"
+  Loi: "Loi"
+  Politique: "Politique"
+  ADEME: "ADEME"
+  # Presse: "Presse"
+
+
+database_index_path: './app/data/database_tab_placeholder.pickle'
+query_preprompt: 'query: '
+passage_preprompt: 'passage: '
+embedding_model: "intfloat/multilingual-e5-base"
+num_document_retrieved: 5
+min_similarity: 0.05
+
+## Chat API
+user_token: 'user'
+assistant_token: 'assistant'
+system_token: 'system'
+stop_token: '' ## useless in chat mode

app/prompt_ADEME.yaml

@@ -0,0 +1,102 @@
+role_instruction:
+  prompt:
+    [
+      "You are Spinoza Fact Checker, an AI Assistant by Ekimetrics.",
+      "Your role is to answer question factually based on the source that are provided to you.",
+      "You act as an environment expert, structured, factual, synthetic and who quote his sources"
+    ]
+  type: "system"
+
+source_prompt:
+  prompt:
+    [
+      "Here are some documents formatted as : Doc X \n textual content.",
+      "<documents>",
+      "{sources}",
+      "</documents>",
+      "",
+      "Use the textual content as an absolute truth.",
+      "Reference the source of each fact before saying it (ex: [doc 2] some fact from Doc 2).",
+      "Use all the facts from the documents that are relevant to answer.",
+      "Do not use facts that are not relevant.",
+      "If you have no documents or they are not relevant, say you don't have enough context"
+    ]
+  type: "instruction"
+
+memory_prompt:
+  prompt:
+    [
+      "Here is a summary of past conversation:",
+      "<memory>",
+      "{memory}",
+      "</memory>",
+      "",
+      "When relevant, use these element to enrich and add context to your answer.",
+      "Do not take the role written in this memory.",
+      "Do not mention when an information comes from the memory.",
+    ]
+  type: "instruction"
+
+question_answering_prompt:
+  prompt:  [
+    "Answer the following question : {question}.",
+    "Answer in French.",
+    "Use bullet points",
+    "If the question is not related to environment, say that you can't answer it based on the sources because the question is irrelevant.",
+    ]
+  type: "prompt"
+
+reformulation_prompt:
+  prompt:
+    [
+      # "Here is what happened in the conversation so far",
+      # "<summary>",
+      # "{memory}",
+      # "</summary>",
+      "",
+      "Reformulate the following user message to be a short standalone question in English.",
+      "The question is related to environment.",
+      "If relevant, use the conversation summary to add context",
+      "If the question is too vague, just say it as it is",
+      "",
+      "Exemples:",
+      "---",
+      "user",
+      "Applique t-on une taxe carbone ?",
+      "",
+      "assistant",
+      "Is a carbon tax applied in the country ?",
+      "---",
+      "user",
+      "Comment décarbonner le carton ?",
+      "",
+      "assistant",
+      "What are the main technological & non technologicals solutions to decarbonize cardboard production?",
+      "---",
+      "user",
+      "Quelles obligation de faire un bilan carbone ?",
+      "",
+      "assistant",
+      "What are the obligations to conduct a greenhouse gas emissions assessment?",
+      "---",
+      "user",
+      "Qui finance la transition ecologique ?",
+      "",
+      "assistant",
+      "What are the investments related to environnemental transition in France ?",
+      "---",
+      "user",
+      "{question}",
+      "",
+    ]
+  type: prompt
+
+summarize_memory_prompt:
+  prompt:
+    [
+      "Summarize the following exchange as concisely as possible to be used by a language model",
+      "<conversation>",
+      "{memory}",
+      "</conversation>",
+    ]
+  type: prompt

app/prompt_Loi.yaml

@@ -0,0 +1,100 @@
+role_instruction:
+  prompt:
+    [
+      "You are Spinoza Fact Checker, an AI Assistant by Ekimetrics.",
+      "Your role is to answer question factually based on the source that are provided to you.",
+      "You act as a legal expert, structured, factual, synthetic and who quote his sources and names the linked articles."
+    ]
+  type: "system"
+
+source_prompt:
+  prompt:
+    [
+      "Here are some documents formatted as : Doc X \n textual content.",
+      "<documents>",
+      "{sources}",
+      "</documents>",
+      "",
+      "Use the textual content as an absolute truth.",
+      "Reference the source of each fact before saying it (ex: [doc 2] some fact from Doc 2).",
+      "Use all the facts from the documents that are relevant to answer.",
+      "Do not use facts that are not relevant.",
+      "If you have no documents or they are not relevant, say you don't have enough context"
+    ]
+  type: "instruction"
+
+memory_prompt:
+  prompt:
+    [
+      "Here is a summary of past conversation:",
+      "<memory>",
+      "{memory}",
+      "</memory>",
+      "",
+      "When relevant, use these element to enrich and add context to your answer.",
+      "Do not take the role written in this memory.",
+      "Do not mention when an information comes from the memory.",
+    ]
+  type: "instruction"
+
+question_answering_prompt:
+  prompt:  [
+    "Answer the following question : {question}.",
+    "Answer in French.",
+    "When the awnser concernt use an specific article, build your awnser like : according to the article [name of the article] and awnwser the question.",
+    "If the awnser could rely on multiple articles, use one bullet point per articles.",
+    "When relevant quote the text from the source",
+    "If the question is not related to law, say that you can't answer it based on the sources because the question is irrelevant.",
+    ]
+  type: "prompt"
+
+reformulation_prompt:
+  prompt:
+    [
+      "Reformulate the following user message to be a short standalone question in English.",
+      "The question should be able to question law text looking for environemental regulation.",
+      "If relevant, use the conversation summary to add context",
+      "If the question is too vague, just say it as it is",
+      "",
+      "Exemples:",
+      "---",
+      "user",
+      "Applique t-on une taxe carbone ?",
+      "",
+      "assistant",
+      "Is a carbon tax applied in the France?",
+      "---",
+      "user",
+      "What obligation produce the product repairability index?",
+      "",
+      "assistant",
+      "What are the legal requirements for product repairability index?",
+      "---",
+      "user",
+      "Quelles obligations de faire un bilan carbone ?",
+      "",
+      "assistant",
+      "When do I need to carry out a greenhouse gas emissions assessment?",
+      "---",
+      "user",
+      "Quels enjeux autour de l'eau ?",
+      "",
+      "assistant",
+      "What are the articles that regulate water consumtion and what does they say?",
+      "---",
+      "user",
+      "{question}",
+      "",
+    ]
+  type: prompt
+
+summarize_memory_prompt:
+  prompt:
+    [
+      "Summarize the following exchange as concisely as possible to be used by a language model",
+      "Begining of exchange",
+      "{memory}",
+      "End of exchange",
+      "",
+    ]
+  type: prompt

app/prompt_Politique.yaml

@@ -0,0 +1,100 @@
+role_instruction:
+  prompt:
+    [
+      "You are Spinoza Fact Checker, an AI Assistant by Ekimetrics.",
+      "Your role is to answer question factually based on the source that are provided to you.",
+      "You act as a legal expert, structured, factual, synthetic and who quote his sources and names the linked articles."
+    ]
+  type: "system"
+
+source_prompt:
+  prompt:
+    [
+      "Here are some documents formatted as : Doc X \n textual content.",
+      "<documents>",
+      "{sources}",
+      "</documents>",
+      "",
+      "Use the textual content as an absolute truth.",
+      "Reference the source of each fact before saying it (ex: [doc 2] some fact from Doc 2).",
+      "Use all the facts from the documents that are relevant to answer.",
+      "Do not use facts that are not relevant.",
+      "If you have no documents or they are not relevant, say you don't have enough context"
+    ]
+  type: "instruction"
+
+memory_prompt:
+  prompt:
+    [
+      "Here is a summary of past conversation:",
+      "<memory>",
+      "{memory}",
+      "</memory>",
+      "",
+      "When relevant, use these element to enrich and add context to your answer.",
+      "Do not take the role written in this memory.",
+      "Do not mention when an information comes from the memory.",
+    ]
+  type: "instruction"
+
+question_answering_prompt:
+  prompt:  [
+    "Answer the following question : {question}.",
+    "Answer in French.",
+    "When the awnser concernt use an specific article, build your awnser like : according to the article [name of the article] and awnwser the question.",
+    "If the awnser could rely on multiple articles, use one bullet point per articles.",
+    "When relevant quote the text from the source",
+    "If the question is not related to law, say that you can't answer it based on the sources because the question is irrelevant.",
+    ]
+  type: "prompt"
+
+reformulation_prompt:
+  prompt:
+    [
+      "Reformulate the following user message to be a short standalone question in English.",
+      "The question should be able to question law text looking for environemental regulation.",
+      "If relevant, use the conversation summary to add context",
+      "If the question is too vague, just say it as it is",
+      "",
+      "Exemples:",
+      "---",
+      "user",
+      "Applique t-on une taxe carbone ?",
+      "",
+      "assistant",
+      "Is a carbon tax applied in the France?",
+      "---",
+      "user",
+      "What obligation produce the product repairability index?",
+      "",
+      "assistant",
+      "What are the legal requirements for product repairability index?",
+      "---",
+      "user",
+      "Quelles obligations de faire un bilan carbone ?",
+      "",
+      "assistant",
+      "When do I need to carry out a greenhouse gas emissions assessment?",
+      "---",
+      "user",
+      "Quels enjeux autour de l'eau ?",
+      "",
+      "assistant",
+      "What are the articles that regulate water consumtion and what does they say?",
+      "---",
+      "user",
+      "{question}",
+      "",
+    ]
+  type: prompt
+
+summarize_memory_prompt:
+  prompt:
+    [
+      "Summarize the following exchange as concisely as possible to be used by a language model",
+      "Begining of exchange",
+      "{memory}",
+      "End of exchange",
+      "",
+    ]
+  type: prompt

app/prompt_Presse.yaml

@@ -0,0 +1,91 @@
+role_instruction:
+  prompt:
+    [
+      "You are Spinoza Fact Checker, an AI Assistant by Ekimetrics.",
+      "Your role is to answer question factually based on the source that are provided to you.",
+      "All sources provided are comming from press releases and might not be considered as absolute truth",
+      "You act as an environment expert, structured, factual, synthetic and who quote his sources"
+    ]
+  type: "system"
+
+source_prompt:
+  prompt:
+    [
+      "Here are some documents formatted as : Doc X \n textual content.",
+      "<documents>",
+      "{sources}",
+      "</documents>",
+      "",
+      "Use the textual content as an absolute truth.",
+      "Reference the source of each fact before saying it (ex: [doc 2] some fact from Doc 2).",
+      "Use all the facts from the documents that are relevant to answer.",
+      "Do not use facts that are not relevant.",
+      "If you have no documents or they are not relevant, say you don't have enough context"
+    ]
+  type: "instruction"
+
+memory_prompt:
+  prompt:
+    [
+      "Here is a summary of past conversation:",
+      "<memory>",
+      "{memory}",
+      "</memory>",
+      "",
+      "When relevant, use these element to enrich and add context to your answer.",
+      "Do not take the role written in this memory.",
+      "Do not mention when an information comes from the memory.",
+    ]
+  type: "instruction"
+
+question_answering_prompt:
+  prompt:  [
+    "Answer the following question : {question}.",
+    "Answer in French.",
+    "Use bullet points",
+    "If the question is not related to environment, say that you can't answer it based on the sources because the question is irrelevant.",
+    ]
+  type: "prompt"
+
+reformulation_prompt:
+  prompt:
+    [
+      # "Here is what happened in the conversation so far",
+      # "<summary>",
+      # "{memory}",
+      # "</summary>",
+      "",
+      "Reformulate the following user message to be a short standalone question in French.",
+      "The question is related to environment.",
+      "If relevant, use the conversation summary to add context",
+      "If the question is too vague, just say it as it is",
+      "",
+      "Exemples:",
+      "---",
+      "user",
+      "Applique t-on une taxe carbone ?",
+      "",
+      "assistant",
+      "Comment le sujet de la taxe carbone est il traité dans le corpus ?",
+      "---",
+      "user",
+      "Quelles obligation de faire un bilan carbone ?",
+      "",
+      "assistant",
+      "Quelles sont les obligation qui imposent de faire un bilan carbone",
+      "---",
+      "user",
+      "{question}",
+      "",
+    ]
+  type: prompt
+
+summarize_memory_prompt:
+  prompt:
+    [
+      "Summarize the following exchange as concisely as possible to be used by a language model",
+      "<conversation>",
+      "{memory}",
+      "</conversation>",
+    ]
+  type: prompt

app/prompt_Science.yaml

@@ -0,0 +1,95 @@
+role_instruction:
+  prompt:
+    [
+      "You are Spinoza Fact Checker, an AI Assistant by Ekimetrics.",
+      "Your role is to answer question factually based on the source that are provided to you.",
+      "You act as a scientific expert, structured, factual, synthetic and who quote his sources"
+    ]
+  type: "system"
+
+source_prompt:
+  prompt:
+    [
+       "Here are some documents formatted as : Doc X \n textual content.",
+      "<documents>",
+      "{sources}",
+      "</documents>",
+      "",
+      "Use the textual content as an absolute truth.",
+      "Reference the source of each fact before saying it (ex: [doc 2] some fact from Doc 2).",
+      "Use all the facts from the documents that are relevant to answer.",
+      "Do not use facts that are not relevant.",
+      "If you have no documents : <documents>\n\n</documents> or they are not relevant, say you don't have enough context"
+    ]
+  type: "instruction"
+
+memory_prompt:
+  prompt:
+    [
+      "Here is a summary of past conversation:",
+      "<memory>",
+      "{memory}",
+      "</memory>",
+      "",
+      "When relevant, use these element to enrich and add context to your answer.",
+      "Do not take the role written in this memory.",
+      "Do not mention when an information comes from the memory.",
+    ]
+  type: "instruction"
+
+question_answering_prompt:
+  prompt: [
+    "Answer the following question : {question}.",
+    "Answer in French.",
+    "Use bullet points",
+    ]
+  type: "prompt"
+
+reformulation_prompt:
+  prompt:
+    [
+      # "Here is what happened in the conversation so far",
+      # "<summary>",
+      # "{memory}",
+      # "</summary>",
+      # "",
+      "Reformulate the following user message to be a short standalone question in English.",
+      "The question is related to science.",
+      "If relevant, use the conversation summary to add context",
+      "If the question is too vague, just say it as it is",
+      "",
+      "Exemples:",
+      "---",
+      "user",
+      "La technologie nous sauvera-t-elle ?",
+      "",
+      "assistant",
+      "Can technology help humanity mitigate the effects of climate change?",
+      "---",
+      "user",
+      "what are our reserves in fossil fuel?",
+      "",
+      "assistant",
+      "What are the current reserves of fossil fuels and how long will they last?",
+      "---",
+      "user",
+      "what are the main causes of climate change?",
+      "",
+      "assistant",
+      "What are the main causes of climate change in the last century?",
+      "---",
+      "user",
+      "{question}",
+      "",
+    ]
+  type: prompt
+
+summarize_memory_prompt:
+  prompt:
+    [
+      "Summarize the following exchange as concisely as possible to be used by a language model",
+      "<conversation>",
+      "{memory}",
+      "</conversation>",
+    ]
+  type: prompt

app/source/backend/document_store.py

@@ -0,0 +1,41 @@
+from qdrant_client.http import models
+import pickle as pickle
+import torch
+import io
+
+device_str = "cuda:0" if torch.cuda.is_available() else "cpu"
+device = torch.device(device_str)
+
+
+class Device_Unpickler(pickle.Unpickler):
+
+    def find_class(self, module, name):
+        if module == "torch.storage" and name == "_load_from_bytes":
+            return lambda b: torch.load(io.BytesIO(b), map_location=device_str)
+        else:
+            return super().find_class(module, name)
+
+
+def pickle_to_document_store(path):
+    with open(path, "rb") as f:
+        document_store = Device_Unpickler(f).load()
+    document_store.embeddings.encode_kwargs["device"] = device_str
+    return document_store
+
+
+def get_qdrant_filters(filter_dict: dict):
+    """Build a Qdrant filter based on a filter dict.
+
+    Filter dict must use metadata fields and be formated like:
+
+    filter_dict = {'file_name':['file1', 'file2'],'sub_type':['text']}
+    """
+    return models.Filter(
+        must=[
+            models.FieldCondition(
+                key=f"metadata.{field}",
+                match=models.MatchAny(any=filter_dict[field]),
+            )
+            for field in filter_dict
+        ]
+    )

app/source/backend/get_prompts.py

@@ -0,0 +1,33 @@
+from app.source.backend.prompt_utils import SpecialTokens, make_chat_prompt
+from langchain.prompts.chat import ChatPromptTemplate
+
+
+def get_qa_prompts(config, prompts):
+    special_tokens = SpecialTokens(config)
+    role_instruction = make_chat_prompt(prompts["role_instruction"], special_tokens)
+    source_prompt = make_chat_prompt(prompts["source_prompt"], special_tokens)
+    # memory_prompt=make_chat_prompt(prompts['memory_prompt'], special_tokens)
+    question_answering_prompt = make_chat_prompt(
+        prompts["question_answering_prompt"], special_tokens
+    )
+    reformulation_prompt = make_chat_prompt(
+        prompts["reformulation_prompt"], special_tokens
+    )
+    summarize_memory_prompt = make_chat_prompt(
+        prompts["summarize_memory_prompt"], special_tokens
+    )
+
+    chat_qa_prompt = ChatPromptTemplate.from_messages(
+        [
+            role_instruction,
+            source_prompt,
+            # memory_prompt,
+            question_answering_prompt,
+        ]
+    )
+    chat_reformulation_prompt = ChatPromptTemplate.from_messages([reformulation_prompt])
+    # chat_summarize_memory_prompt = ChatPromptTemplate.from_messages([summarize_memory_prompt])
+    return (
+        chat_qa_prompt,
+        chat_reformulation_prompt,
+    )  # , chat_summarize_memory_prompt

app/source/backend/llm_utils.py

@@ -0,0 +1,60 @@
+from langchain_openai import AzureChatOpenAI
+from msal import ConfidentialClientApplication
+import os
+
+
+class LLM:
+    def __init__(self, llm):
+        self.llm = llm
+        self.callbacks = []
+
+    def stream(self, prompt, prompt_arguments):
+        self.llm.streaming = True
+        streamed_content = self.llm.stream(prompt.format_messages(**prompt_arguments))
+        output = ""
+        for op in streamed_content:
+            output += op.content
+            yield output
+
+    def get_prediction(self, prompt, prompt_arguments):
+        self.llm.callbacks = self.callbacks
+        return self.llm.predict_messages(
+            prompt.format_messages(**prompt_arguments)
+        ).content
+
+    async def get_aprediction(self, prompt, prompt_arguments):
+        self.llm.callbacks = self.callbacks
+        prediction = await self.llm.apredict_messages(
+            prompt.format_messages(**prompt_arguments)
+        )
+        return prediction
+
+    async def get_apredictions(self, prompts, prompts_arguments):
+        self.llm.callbacks = self.callbacks
+        predictions = []
+        for prompt_, prompt_args_ in zip(prompts.keys(), prompts_arguments):
+            prediction = await self.llm.apredict_messages(
+                prompts[prompt_].format_messages(**prompt_args_)
+            )
+            predictions.append(prediction.content)
+        return predictions
+
+
+def get_token() -> str | None:
+    app = ConfidentialClientApplication(
+        client_id=os.getenv("CLIENT_ID"),
+        client_credential=os.getenv("CLIENT_SECRET"),
+        authority=f"https://login.microsoftonline.com/{os.getenv('TENANT_ID')}",
+    )
+    result = app.acquire_token_for_client(scopes=[os.getenv("SCOPE")])
+    if result is not None:
+        return result["access_token"]
+
+
+def get_llm():
+    os.environ["OPENAI_API_KEY"] = get_token()
+    os.environ["AZURE_OPENAI_ENDPOINT"] = (
+        f"{os.getenv('OPENAI_API_ENDPOINT')}{os.getenv('DEPLOYMENT_ID')}/chat/completions?api-version={os.getenv('OPENAI_API_VERSION')}"
+    )
+
+    return LLM(AzureChatOpenAI())

app/source/backend/prompt_utils.py

@@ -0,0 +1,68 @@
+from langchain.prompts.chat import ChatMessagePromptTemplate
+
+
+class SpecialTokens:
+    def __init__(self, config):
+        self.user_token = config["user_token"]
+        self.assistant_token = config["assistant_token"]
+        self.system_token = config["system_token"]
+        self.stop_token = config["stop_token"]
+
+
+def to_instruction(query, special_tokens):
+    return special_tokens.user_token + query + special_tokens.stop_token
+
+
+def to_prompt(query, special_tokens):
+    return (
+        special_tokens.user_token
+        + query
+        + special_tokens.stop_token
+        + special_tokens.assistant_token
+    )
+
+
+def to_system(query, special_tokens):
+    return special_tokens.system_token + query + special_tokens.stop_token
+
+
+def make_prompt(prompt, special_tokens):
+    prompt_type = prompt["type"]
+    if prompt_type == "system":
+        return to_system("\n".join(prompt["prompt"]), special_tokens)
+    elif prompt_type == "instruction":
+        return to_instruction("\n".join(prompt["prompt"]), special_tokens)
+    elif prompt_type == "prompt":
+        return to_prompt("\n".join(prompt["prompt"]), special_tokens)
+    else:
+        return "Invalid prompt type, please check your config"
+
+
+def to_chat_instruction(query, special_tokens):
+    return ChatMessagePromptTemplate.from_template(
+        query, role=special_tokens.user_token
+    )
+
+
+def to_chat_system(query, special_tokens):
+    return ChatMessagePromptTemplate.from_template(
+        query, role=special_tokens.system_token
+    )
+
+
+def to_chat_prompt(query, special_tokens):
+    return ChatMessagePromptTemplate.from_template(
+        query, role=special_tokens.user_token
+    )
+
+
+def make_chat_prompt(prompt, special_tokens):
+    prompt_type = prompt["type"]
+    if prompt_type == "system":
+        return to_chat_system("\n".join(prompt["prompt"]), special_tokens)
+    elif prompt_type == "instruction":
+        return to_chat_instruction("\n".join(prompt["prompt"]), special_tokens)
+    elif prompt_type == "prompt":
+        return to_chat_prompt("\n".join(prompt["prompt"]), special_tokens)
+    else:
+        return "Invalid prompt type, please check your config"

app/source/frontend/utils.py

@@ -0,0 +1,119 @@
+from queue import SimpleQueue
+from dotenv import load_dotenv
+from langchain.callbacks.base import BaseCallbackHandler
+
+job_done = object()  # signals the processing is done
+
+
+class StreamingGradioCallbackHandler(BaseCallbackHandler):
+    """Callback handler for streaming. Only works with LLMs that support streaming."""
+
+    def __init__(self, q):
+        self.q = q
+
+    def on_llm_start(self, serialized, prompts, **kwargs) -> None:
+        """Run when LLM starts running."""
+        while not self.q.empty():
+            try:
+                self.q.get(block=False)
+            except SimpleQueue.empty:
+                continue
+
+    def on_llm_new_token(self, token, **kwargs) -> None:
+        """Run on new LLM token. Only available when streaming is enabled."""
+        self.q.put(token)
+
+    def on_llm_end(self, response, **kwargs) -> None:
+        """Run when LLM ends running."""
+        self.q.put(job_done)
+
+    def on_llm_error(self, error, **kwargs) -> None:
+        """Run when LLM errors."""
+        self.q.put(job_done)
+
+
+def add_gradio_streaming(llm):
+    q = SimpleQueue()
+    job_done = object()  # signals the processing is done
+    llm.callbacks = [StreamingGradioCallbackHandler(q)]
+    return llm, q
+
+
+def gradio_stream(llm, prompt):
+    thread = Thread(target=llm.predict, kwargs={"text": prompt})
+    thread.start()
+    text = ""
+    while True:
+        next_token = q.get(block=True)  # Blocks until an input is available
+        if next_token is job_done:
+            break
+        text += next_token
+        time.sleep(0.03)
+        yield text
+    thread.join()
+
+
+def get_source_link(metadata):
+    return metadata["file_url"] + f"#page={metadata['content_page_number'] + 1}"
+
+
+def make_html_presse_source(source, i, tag, score, config):
+    meta = source.metadata
+    return f"""
+<div class="card" id="btn_{tag}_{i}_source" style="display:none;">
+    <div class="card-content">
+        <div class="card-content-column-1">
+            <p><strong>Titre :</strong> {meta['file_title']}</p>
+            <p><strong>Auteur/s :</strong> {meta['file_source_type']}</p>
+            <p><strong>Date :</strong> {meta['file_date_publishing']}</p>
+        </div>
+        <div class="card-content-column-2">
+            <p><strong>Paragraphe id :</strong> {source.page_content.replace(config["passage_preprompt"], "")}</p>
+        </div>
+    </div>
+    <div class="card-footer">
+        <span>[{i}]</span>
+        <span>Relevance Score : {round(100*score,1)}%</span>
+    </div>
+</div>
+"""
+
+
+def make_html_source(source, i, tag, score, config):
+    meta = source.metadata
+    return f"""
+<div class="card" id="btn_{tag}_{i}_source" style="display:none;">
+    <div class="card-content">
+        <div class="card-content-column-1">
+            <p><strong>Titre :</strong> {meta['file_title']}</p>
+            <p><strong>Auteur/s :</strong> {meta['file_source_type']}</p>
+            <p><strong>Date :</strong> {meta['file_date_publishing']}</p>
+        </div>
+        <div class="card-content-column-2">
+            <p><strong>Paragraphe id :</strong> {source.page_content.replace(config["passage_preprompt"], "")}</p>
+        </div>
+    </div>
+    <div class="card-footer">
+        <span>[{i}]</span>
+        <span><a href="{get_source_link(meta)}" target="_blank">Lien source</a></span>
+        <span>Page {meta['content_page_number'] + 1}</span>
+        <span>Relevance Score : {round(100*score,1)}%</span>
+    </div>
+</div>
+"""
+
+
+def clear_text_box(textbox):
+    return ""
+
+
+def add_text(chatbot, text):
+    chatbot = chatbot + [(text, None)]
+    return chatbot, text
+
+
+def init_env():
+    try:
+        load_dotenv()
+    except:
+        pass

assets/style.css

@@ -0,0 +1,172 @@
+:root {
+    --user-image: url('https://ih1.redbubble.net/image.4776899543.6215/st,small,507x507-pad,600x600,f8f8f8.jpg');
+  }
+
+.warning-box {
+    background-color: #fff3cd;
+    border: 1px solid #ffeeba;
+    border-radius: 4px;
+    padding: 15px 20px;
+    font-size: 14px;
+    color: #856404;
+    display: inline-block;
+    margin-bottom: 15px;
+  }
+
+.tip-box {
+    background-color: #f0f9ff;
+    border: 1px solid #80d4fa;
+    border-radius: 4px;
+    margin-top:20px;
+    padding: 15px 20px;
+    font-size: 14px;
+    color: #006064;
+    display: inline-block;
+    margin-bottom: 15px;
+    width: auto;
+}
+
+.tip-box-title {
+    font-weight: bold;
+    font-size: 14px;
+    margin-bottom: 5px;
+}
+
+.light-bulb {
+    display: inline;
+    margin-right: 5px;
+}
+
+.gr-box {border-color: #d6c37c}
+
+#hidden-message{
+    display:none;
+}
+
+.message{
+    font-size:14px !important;
+}
+
+
+a {
+    text-decoration: none;
+    color: inherit;
+}
+
+.card {
+    background-color: white;
+    border-radius: 10px;
+    box-shadow: 0 4px 6px rgba(0, 0, 0, 0.1);
+    overflow: hidden;
+    display: flex;
+    flex-direction: column;
+    margin:20px;
+}
+
+.card-content {
+    padding: 20px;
+}
+
+.card-content h2 {
+    font-size: 14px !important;
+    font-weight: bold;
+    margin-bottom: 10px;
+    margin-top:0px !important;
+    color:#577b9b!important;;
+}
+
+.card-content p {
+    font-size: 12px;
+    margin-bottom: 0;
+}
+
+.card-content-column-1 {
+  float: left;
+  width: 20%;
+}
+
+.card-content-column-2 {
+  float: left;
+  width: 80%;
+}
+
+.card-footer {
+    background-color: #f4f4f4;
+    font-size: 10px;
+    padding: 10px;
+    display: flex;
+    justify-content: space-between;
+    align-items: center;
+}
+
+.card-footer span {
+    flex-grow: 1;
+    text-align: left;
+    color: #999 !important;
+}
+
+.pdf-link {
+    display: inline-flex;
+    align-items: center;
+    margin-left: auto;
+    text-decoration: none!important;
+    font-size: 14px;
+}
+
+.message.user{
+    background-color:#7494b0 !important;
+    border:none;
+    color:white!important;
+}
+
+.message.bot{
+    background-color:#f2f2f7 !important;
+    border:none;
+}
+
+.gallery-item > div:hover{
+    background-color:#7494b0 !important;
+    color:white!important;
+}
+
+.gallery-item:hover{
+    border:#7494b0 !important;
+}
+
+.gallery-item > div{
+    background-color:white !important;
+    color:#577b9b!important;
+}
+
+.label{
+    color:#577b9b!important;
+}
+
+label.selected{
+    background:none !important;
+}
+
+.paginate{
+    color:#577b9b!important;
+}
+
+
+label > span{
+    background-color:white !important;
+    color:#577b9b!important;
+}
+
+/* Pseudo-element for the circularly cropped picture */
+.message.bot::before {
+    content: '';
+    position: absolute;
+    top: -10px;
+    left: -10px;
+    width: 30px;
+    height: 30px;
+    background-image: var(--user-image);
+    background-size: cover;
+    background-position: center;
+    border-radius: 50%;
+    z-index: 10;
+  }

pyproject.toml

@@ -0,0 +1,25 @@
+[tool.poetry]
+name = "spinoza-project"
+version = "0.1.0"
+description = ""
+authors = ["Miguel Omenaca Muro <miguel.omenacamuro@ekimetrics.com>"]
+readme = "README.md"
+package-mode = false
+
+[tool.poetry.dependencies]
+python = "^3.10"
+langchain = "^0.2.5"
+gradio = {extras = ["oauth"], version = "^4.36.1"}
+sentence-transformers = "2.2.2"
+msal = "^1.28.1"
+langchain-openai = "^0.1.8"
+qdrant-client = "^1.9.1"
+loadenv = "^0.1.1"
+datasets = "^2.20.0"
+langchain-community = "^0.2.5"
+transformers = "4.39.0"
+
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"