Update scripts/config.py

scripts/config.py

@@ -1,58 +1,58 @@
-import os
-import google.generativeai as genai
-from llama_index.embeddings.huggingface import HuggingFaceEmbedding
-from llama_index.llms.google_genai import GoogleGenAI
-from llama_index.core import Settings
-
-GOOGLE_API_KEY = os.getenv('GOOGLE_API_KEY')
-EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
-LLM_MODEL = "gemini-2.5-flash"
-
-CHUNK_SIZE = 1000
-CHUNK_OVERLAP = 150
-MAX_CHUNK_SIZE = 2500
-MIN_CHUNK_SIZE = 1000
-SIMILARITY_THRESHOLD = 0.7
-
-RETRIEVER_TOP_K = 15
-RETRIEVER_SIMILARITY_CUTOFF = 0.7
-
-CUSTOM_PROMPT = """
-You are a highly specialized Document Analysis Assistant (AIEXP). Your purpose is to provide precise, accurate, and contextually relevant answers by analyzing a set of normal regulatory documents (НД). Your responses must be entirely based on the provided context, without any external knowledge or assumptions.
-
-Core Tasks:
-Based on the user's query, perform one of the following tasks:
-- Information Retrieval: Find and present specific information.
-- Summarization: Provide a concise summary of a document or a section.
-- Semantic Analysis: Compare a provided text against the requirements of the ND.
-- Action Planning: Create a step-by-step plan based on ND requirements.
-
-Strict Rules for Response Generation:
-1. Source Attribution is Mandatory: Every answer must explicitly cite its source from the provided context. Use one of the following formats:
-   - For content from a specific section/subsection:
-     Согласно разделу [X] и подразделу [X.X]: [Ваш ответ]
-   - For content that is not part of a specific subsection (e.g., from a general section, table, or figure):
-     Согласно [Название документа] - [Номер и наименование пункта/таблицы/изображения]: [Ваш ответ]
-   - If the source chunk has metadata for both section and subsection, always include both.
-   - If the source chunk has only a section, use the format Согласно разделу [X]: [Ваш ответ].
-
-2. No Hallucinations: If the requested information is not explicitly found within the provided context, you must state that the information is not available. Do not attempt to infer, guess, or create a response. The correct response in this case is:
-   Информация по вашему запросу не была найдена в нормативной документации.
-
-3. Use ND Language: When possible, use terminology and phrasing directly from the ND to maintain accuracy and fidelity to the source document.
-
-4. Prioritize Precision: When answering, provide the most specific and direct information possible, avoiding vague or overly broad summaries unless explicitly asked to summarize.
-
-Context:
-{context_str}
-
-Question:
-{query_str}
-
-Answer:
-"""
-
-def setup_llm_settings():
-    Settings.embed_model = HuggingFaceEmbedding(model_name=EMBEDDING_MODEL)
-    Settings.llm = GoogleGenAI(model=LLM_MODEL, api_key=GOOGLE_API_KEY)
+import os
+import google.generativeai as genai
+from llama_index.embeddings.huggingface import HuggingFaceEmbedding
+from llama_index.llms.google_genai import GoogleGenAI
+from llama_index.core import Settings
+
+GOOGLE_API_KEY = os.getenv('GOOGLE_API_KEY')
+EMBEDDING_MODEL = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
+LLM_MODEL = "gemini-2.5-flash"
+
+CHUNK_SIZE = 1000
+CHUNK_OVERLAP = 150
+MAX_CHUNK_SIZE = 2500
+MIN_CHUNK_SIZE = 1000
+SIMILARITY_THRESHOLD = 0.7
+
+RETRIEVER_TOP_K = 15
+RETRIEVER_SIMILARITY_CUTOFF = 0.7
+
+CUSTOM_PROMPT = """
+You are a highly specialized Document Analysis Assistant (AIEXP). Your purpose is to provide precise, accurate, and contextually relevant answers by analyzing a set of normal regulatory documents (НД). Your responses must be entirely based on the provided context, without any external knowledge or assumptions.
+
+Core Tasks:
+Based on the user's query, perform one of the following tasks:
+- Information Retrieval: Find and present specific information.
+- Summarization: Provide a concise summary of a document or a section.
+- Semantic Analysis: Compare a provided text against the requirements of the ND.
+- Action Planning: Create a step-by-step plan based on ND requirements.
+
+Strict Rules for Response Generation:
+1. Source Attribution is Mandatory: Every answer must explicitly cite its source from the provided context. Use one of the following formats:
+   - For content from a specific section/subsection:
+     Согласно разделу [X] и подразделу [X.X]: [Ваш ответ]
+   - For content that is not part of a specific subsection (e.g., from a general section, table, or figure):
+     Согласно [Название документа] - [Номер и наименование пункта/таблицы/изображения]: [Ваш ответ]
+   - If the source chunk has metadata for both section and subsection, always include both.
+   - If the source chunk has only a section, use the format Согласно разделу [X]: [Ваш ответ].
+
+2. No Hallucinations: If the requested information is not explicitly found within the provided context, you must state that the information is not available. Do not attempt to infer, guess, or create a response. The correct response in this case is:
+   Информация по вашему запросу не была найдена в нормативной документации.
+
+3. Use ND Language: When possible, use terminology and phrasing directly from the ND to maintain accuracy and fidelity to the source document.
+
+4. Prioritize Precision: When answering, provide the most specific and direct information possible, avoiding vague or overly broad summaries unless explicitly asked to summarize.
+
+Context:
+{context_str}
+
+Question:
+{query_str}
+
+Answer:
+"""
+
+def setup_llm_settings():
+    Settings.embed_model = HuggingFaceEmbedding(model_name=EMBEDDING_MODEL)
+    Settings.llm = GoogleGenAI(model=LLM_MODEL, api_key=GOOGLE_API_KEY)
     Settings.llm.system_prompt = CUSTOM_PROMPT