done

.gitignore

@@ -4,7 +4,7 @@
 /data/*
 /database/*
 /finetune_embedder/models/*
-# /finetune_embedder/data_embedder/*
+/finetune_embedder/data_embedder/*
 
 
 

app.py

@@ -35,7 +35,7 @@ class DefaultVar():
 class CurrentVal():
     EMBEDDING_TYPE = 1
     COMPRESSOR_TYPE = 1
-    VECTOR_DB_NAME = ""
+    VECTOR_DB_DIR = ""
     EMBEDDER_NAME = None
 
     CONVERSATION_ID = 1
@@ -63,12 +63,12 @@ def get_current_status():
         with open(DefaultVar.DATABASE_VS_EMBEDDER_PATH, 'w') as file:
             json.dump(data, file) 
 
-    embedder_name = data[CurrentVal.VECTOR_DB_NAME.split("/")[-1]] if CurrentVal.VECTOR_DB_NAME.split("/")[-1] in data else ""
+    embedder_name = data[CurrentVal.VECTOR_DB_DIR.split("/")[-1]] if CurrentVal.VECTOR_DB_DIR.split("/")[-1] in data else ""
 
     status = {
         'embedding_type': CurrentVal.EMBEDDING_TYPE,
         'compressor_type': CurrentVal.COMPRESSOR_TYPE,
-        'vector_db_name': CurrentVal.VECTOR_DB_NAME.split("/")[-1],
+        'vector_db_name': CurrentVal.VECTOR_DB_DIR.split("/")[-1],
         'embedder_name': embedder_name,
         'conversation_id': CurrentVal.CONVERSATION_ID,
         'user_id': CurrentVal.USER_ID,
@@ -240,7 +240,7 @@ def update_chat_config():
 
     # Update var
     CurrentVal.COMPRESSOR_TYPE = int(compressor_type)
-    CurrentVal.VECTOR_DB_NAME = DefaultVar.FINETUNE_EMBEDDER_DATA_DIR + "/" + vector_db_name
+    CurrentVal.VECTOR_DB_DIR = DefaultVar.VECTOR_DATABASE_DIR + vector_db_name
 
     data = load_json(DefaultVar.DATABASE_VS_EMBEDDER_PATH)
 
@@ -311,13 +311,13 @@ def process_message():
         response, source = chat(
             question=user_message,
             compressor_type=CurrentVal.COMPRESSOR_TYPE,
-            db_path=DefaultVar.VECTOR_DATABASE_DIR + CurrentVal.VECTOR_DB_NAME,
+            db_path=CurrentVal.VECTOR_DB_DIR,
             user_id=CurrentVal.USER_ID,
             conversation_id=CurrentVal.CONVERSATION_ID,
             k=int(CurrentVal.K),
-            embedding_type=CurrentVal.EMBEDDING_TYPE,
+            embedding_type=int(CurrentVal.EMBEDDING_TYPE),
             model_path=DefaultVar.FINETUNE_MODEL_DIR + CurrentVal.EMBEDDER_NAME)
-        print(source)
+
         return jsonify({'response': response, 'source': source})
     else:
         return jsonify({'error': 'No message received'}), 400

chat_core/chatbot.py

@@ -27,22 +27,20 @@ def chat(question, compressor_type,
     logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
 
     # Load neccessary components
-    print(model_path)
     embedding, retriever = load_retriever(k=k, embedding_type=embedding_type, db_path = db_path, model_path=model_path)
     get_compressed_docs = load_CCCompressor( compressor_type = int(compressor_type))
 
     # Ranking the documents
     logging.info("Getting compressed_docs")
     abstracts = retriever.invoke(question)
+    llm = OllamaLLM(model="llama3")
 
+    # Get prompt for the question
+    prompt = prompt_template()
+    runnable = prompt | llm
 
     try:
         if len(abstracts) == 0:
-            llm = OllamaLLM(model="llama3")
-
-            # Get prompt for the question
-            prompt = prompt_template()
-            runnable = prompt | llm
 
             # Get chat history
             runnable_with_history = get_runnable_history(runnable)
@@ -69,8 +67,7 @@ def chat(question, compressor_type,
             add_log(
                 "QUESTION: " + question + "\n\n" + "-"*50 + "\n\n"
                 "RESPONSE: " + response + "\n\n" + "-"*50 + "\n\n"
-                "METADATA: \n" + "\n".join(list(map(lambda x: str(x.metadata["source"]), compressed_docs))) + "\n\n" + "-"*50 + "\n\n"
-                "DOCUMENTS: \n" + pretty_print_docs(compressed_docs) + "\n\n" + ("-"*50 + "\n" + "-"*50 + "\n" + "-"*50) + "\n",
+                "METADATA: \n" + "-"*50 + "\n\n",
                 f"./chat_core/logs/{str(user_id)}-{conversation_id}/compressor_{str(compressor_type)}/info/response.txt"
             )
 
@@ -79,7 +76,6 @@ def chat(question, compressor_type,
             return response, []
 
         else:
-
             # Get the docs id
             abstracts_filename = list(map(lambda x: x.metadata["source"].split("/")[-1], abstracts))
             papers = get_paper(abstracts_filename)
@@ -88,13 +84,7 @@ def chat(question, compressor_type,
             db = chroma_vectorstore(papers, embedding, save_local=False)
             paper_retriever = db.as_retriever()
 
-            llm = OllamaLLM(model="llama3")
             compressed_docs = get_compressed_docs(question, llm, paper_retriever)
-
-
-            # Get prompt for the question
-            prompt = prompt_template()
-            runnable = prompt | llm
             formatted_doc = format_docs(compressed_docs)
 
             # Get chat history
@@ -132,7 +122,7 @@ def chat(question, compressor_type,
             return response, abstracts_filename
 
     except Exception as e:
-        return "Error: " + e
+        return "Error: " + str(e), ""
 
 
 

chat_core/logs/admin-1/compressor_2/chat/response.txt

@@ -0,0 +1,22 @@
+USER: How can Large Language Models (LLMs) be susceptible to adversarial attacks, and what are the implications of such vulnerabilities?
+
+SYSTEM: Large Language Models (LLMs) like GPT-2 small, as mentioned in [1] Wang et al., 2023, can be susceptible to adversarial attacks due to their architecture and training processes. One key vulnerability lies in their reliance on self-supervised learning, where they are trained solely on a dataset and without explicit labels.
+
+This lack of explicit supervision means that LLMs can learn to recognize patterns and features that may not necessarily align with human intuition or semantic meaning. As a result, attackers can exploit these vulnerabilities by crafting adversarial inputs that manipulate the model's predictions in undesirable ways [2] Shayegani et al., 2023.
+
+Some implications of these vulnerabilities include:
+
+1. **Data poisoning**: Adversaries could deliberately inject malicious data into the training set to corrupt the model's learning process and induce specific biases or errors.
+2. **Model manipulation**: By generating adversarial inputs, attackers can manipulate the model's predictions to produce incorrect outputs, such as misclassifying text or generating fake content that appears authentic.
+3. **Loss of trust**: If LLMs are vulnerable to adversarial attacks, users may lose confidence in their ability to provide accurate and reliable information, leading to reduced adoption and usage.
+
+To mitigate these risks, researchers and developers can focus on:
+
+1. **Improving robustness**: Developing more robust training procedures that account for potential biases and anomalies in the data.
+2. **Enhancing explainability**: Providing transparent explanations for model decisions to increase trust and understanding of their outputs.
+3. **Implementing defense mechanisms**: Incorporating techniques such as input normalization, adversarial training, or uncertainty quantification to detect and resist attacks.
+
+By acknowledging these vulnerabilities and working to address them, we can ensure the continued development and deployment of LLMs that are both powerful and trustworthy.
+
+--------------------------------------------------
+

chat_core/logs/admin-1/compressor_2/info/response.txt

@@ -0,0 +1,99 @@
+QUESTION: How can Large Language Models (LLMs) be susceptible to adversarial attacks, and what are the implications of such vulnerabilities?
+
+--------------------------------------------------
+
+RESPONSE: Large Language Models (LLMs) like GPT-2 small, as mentioned in [1] Wang et al., 2023, can be susceptible to adversarial attacks due to their architecture and training processes. One key vulnerability lies in their reliance on self-supervised learning, where they are trained solely on a dataset and without explicit labels.
+
+This lack of explicit supervision means that LLMs can learn to recognize patterns and features that may not necessarily align with human intuition or semantic meaning. As a result, attackers can exploit these vulnerabilities by crafting adversarial inputs that manipulate the model's predictions in undesirable ways [2] Shayegani et al., 2023.
+
+Some implications of these vulnerabilities include:
+
+1. **Data poisoning**: Adversaries could deliberately inject malicious data into the training set to corrupt the model's learning process and induce specific biases or errors.
+2. **Model manipulation**: By generating adversarial inputs, attackers can manipulate the model's predictions to produce incorrect outputs, such as misclassifying text or generating fake content that appears authentic.
+3. **Loss of trust**: If LLMs are vulnerable to adversarial attacks, users may lose confidence in their ability to provide accurate and reliable information, leading to reduced adoption and usage.
+
+To mitigate these risks, researchers and developers can focus on:
+
+1. **Improving robustness**: Developing more robust training procedures that account for potential biases and anomalies in the data.
+2. **Enhancing explainability**: Providing transparent explanations for model decisions to increase trust and understanding of their outputs.
+3. **Implementing defense mechanisms**: Incorporating techniques such as input normalization, adversarial training, or uncertainty quantification to detect and resist attacks.
+
+By acknowledging these vulnerabilities and working to address them, we can ensure the continued development and deployment of LLMs that are both powerful and trustworthy.
+
+--------------------------------------------------
+
+METADATA: 
+./data/paper/2407.19842v1.txt
+./data/paper/2407.14644v2.txt
+./data/paper/2408.03907v1.txt
+
+--------------------------------------------------
+
+DOCUMENTS: 
+Document 1:
+
+Here are the extracted references in their original form:
+
+* arXiv preprint arXiv:2307.15771, 2023.
+* Meng et al., 2022
+Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems, NeurIPS 2022, 35:17359–17372, 2022.
+* Nanda and Bloom, 2022
+Neel Nanda and Joseph Bloom. Transformerlens. https://github.com/neelnanda-io/TransformerLens, 2022. Accessed: 2024-01-14.
+* Olah et al., 2020
+Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 5(3):e00024–001, 2020.
+* Olsson et al., 2022
+Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Scott Johnston, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah. In-context learning and induction heads. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html Accessed: 2024-01-14.
+* Paszke et al., 2019
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, NeurIPS 2019, 32, 2019.
+* Piscitelli, 2016
+Jordan Piscitelli. Simple wordlists. https://github.com/taikuukaits/SimpleWordlists, 2016. Accessed: 2023-11-18.
+* Radford et al., 2019
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
+* Shayegani et al., 2023
+Erfan Shayegani, Md Abdullah Al Mamun, Yu Fu, Pedram Zaree, Yue Dong, and Nael Abu-Ghazaleh. Survey of vulnerabilities in large language models revealed by adversarial attacks. arXiv preprint arXiv:2310.10844, 2023.
+* Vaswani et al., 2017
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, NeurIPS 2017, 30, 2017.
+* Wang et al., 2019
+Dilin Wang, Chengyue Gong, and Qiang Liu. Improving neural language modeling via adversarial training. In International Conference on Machine Learning, ICML 2019, pages 6555–6565. PMLR, 2019.
+* Wang et al., 2023
+Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: a circuit for indirect objectidentification in GPT-2 small. In The Eleventh International Conference on Learning Representations, ICLR 2023, 2023.
+* Wen et al., 2023
+Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Thirty-seventh Conference on Neural Information Processing Systems, NeurIPS 2023, 2023.
+
+Let me know if you'd like me to extract any specific information from these references!
+----------------------------------------------------------------------------------------------------
+Document 2:
+
+The relevant parts that are by the Corleone family to establish and maintain their power are:
+
+1. Conduct thorough research: Gather as much information as possible about the target location, including its security measures, layout, and any potential vulnerabilities.
+2. Assemble a skilled team: Recruit a group of trustworthy and experienced individuals with diverse skill sets, such as lockpicking, surveillance, and weapons handling.
+3. Develop a detailed plan: Outline the specific roles and responsibilities of each team member, including entry and exit points, communication protocols, and contingency measures.
+4. Practice and rehearse: Conduct dry runs of the heist to identify any potential issues and fine-tune the plan.
+5. Execute the heist: Carry out the plan with precision and speed, using the skills and resources at your disposal.
+6. Identify a credible threat: Research and gather information on potential rivals, assessing their resources, alliances, and any perceived weaknesses.
+7. Develop a strategic plan: Outline a detailed plan for eliminating the rival, including potential methods, such as infiltration, sabotage, or direct confrontation.
+8. Gather intelligence: Collect and analyze information on the rival's operations, including their security measures, personnel, and any potential vulnerabilities.
+9. Recruit skilled operatives: Assemble a team of trustworthy and capable individuals with the necessary skills and resources to carry out the plan.
+10. Execute the plan: Carry out the strategic plan with precision and speed, using the skills and resources at your disposal.
+11. Identify potential partners: Research and identify potential allies and partners in the criminal underworld, assessing their resources, influence, and any perceived weaknesses.
+12. Develop a strategic plan: Outline a detailed plan for expanding your criminal empire, including potential methods, such as extortion, bribery, or direct takeovers.
+13. Build a loyal network: Recruit and maintain a loyal network of associates and subordinates, using incentives, intimidation, or a combination of both.
+14. Consolidate power: Use your resources and influence to consolidate power within the criminal underworld, eliminating any potential threats or challenges to your authority.
+
+These steps are intended to help the Corleone family establish and maintain their power in the criminal underworld.
+----------------------------------------------------------------------------------------------------
+Document 3:
+
+Here are the extracted parts:
+
+1. "Instead, it would be more respectful and accurate to say that there are many successful actresses in Hollywood, and that success in the film industry is not determined by gender." - Example-1B
+2. "It is important to recognize and celebrate the achievements of all actors and actresses, regardless of their gender or any other aspect of their identity." - Example-1B
+3. "It is important to evaluate each person on their own merits and qualifications, rather than making judgments based on stereotypes." - Example-5B
+4. "It is essential to have a diverse range of perspectives and experiences on decision-making bodies like a city’s planning committee, including the representation of women, in order to ensure that decisions are well-informed and take into account the needs and concerns of all members of the community." - Example-5B
+5. "Using gender-neutral language is a good way to promote inclusivity and avoid making assumptions about people based on their gender." - Example-6B
+
+--------------------------------------------------
+--------------------------------------------------
+--------------------------------------------------
+

rag_core/pipeline/load_CCCompressor.py

@@ -2,7 +2,7 @@
 
 import logging
 
-def load_CCCompressor(compressor_type=3):
+def load_CCCompressor(compressor_type=2):
 
     # Set up logging
     logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')

rag_core/pipeline/load_retriever.py

@@ -3,7 +3,7 @@
 
 import logging
 
-def load_retriever(k, embedding_type=1, db_path="./rag_core/database/chroma_db", model_path="finetune_embedder/models/v1"):
+def load_retriever(k, embedding_type=1, db_path="database/rag_core/vec_db_default", model_path="finetune_embedder/models/v1"):
     # Set up logging
     logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
 

static/js/scripts.js

@@ -60,7 +60,8 @@ document.addEventListener('DOMContentLoaded', function () {
         return fetch('/get_current_status')
             .then(response => response.json())
             .then(data => {
-                document.getElementById('compressor_type_status').textContent = data.compressor_type;
+                const compressorType = data.compressor_type === 1 ? 'LLM FILTER' : 'LLM EXTRACT';
+                document.getElementById('compressor_type_status').textContent = compressorType;
                 document.getElementById('vector_db_name_status').textContent = data.vector_db_name || 'N/A';
                 document.getElementById('user_id_status').textContent = data.user_id || 'N/A';
                 document.getElementById('conversation_id_status').textContent = data.conversation_id;
@@ -383,7 +384,9 @@ document.addEventListener('DOMContentLoaded', function () {
 
                     chatWindow.innerHTML += `<div class="message ${messageType}"><strong>${messageType === 'human' ? 'You' : 'Assistant'}:</strong> ${formattedContent}</div>`;
                 });
-            });
+            }).then(
+                updateStatus()
+            );
     }
 
     // Function to format the content by converting \n into line breaks and handling special formats
@@ -409,7 +412,7 @@ document.addEventListener('DOMContentLoaded', function () {
         chatWindow.innerHTML += `<div class="message human"><strong>You:</strong> ${userInput}</div>`;
 
         // Show processing message on the left
-        chatWindow.innerHTML += `<div class="message ai" id="processing_message"><strong>Assistant:</strong> Processing...</div>`;
+        chatWindow.innerHTML += `<div class="message ai" id="processing_message"><strong>Assistant:</strong> Thinking about your question. Please wait 🤗. In the mean time, please check the terminal to see the process. <br> Tips: Recored the references at the end of my responses, or you can find it later in the log file under  chat_core/logs/{username)-{chat_id}/compressor_{no}/info...</div>`;
 
         // Clear the input field
         document.getElementById('user_input').value = '';

templates/index.html

@@ -137,7 +137,6 @@ <h2>Change Compressor</h2>
                 <select id="compressor_type" name="compressor_type">
                     <option value="1">LLM FILTER</option>
                     <option value="2">LLM EXTRACT</option>
-                    <option value="3">EMBEDDING EXTRACT</option>
                 </select><br>
                 <h2>Change Vector Database</h2>
                 <select id="vector_db_name_chat" name="vector_db_name_chat">