Spaces:
Building
Building
File size: 9,498 Bytes
d606279 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 |
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### **Implementing Retrieval Ranking in Your RAG System (LangChain + ChromaDB)** \n",
"To improve the retrieval quality in your **RAG pipeline**, you can implement **retrieval ranking** before passing the retrieved documents to the LLM. This ensures the **most relevant** documents are prioritized.\n",
"\n",
"---\n",
"\n",
"## **🔹 Why is Retrieval Ranking Important?**\n",
"When using a **vector database like ChromaDB**, similarity search retrieves the `k` closest documents. However, these documents:\n",
"1. **May not always be in the best order** based on query relevance.\n",
"2. **Might contain redundant or less useful information**.\n",
"3. **Need re-ranking to prioritize more relevant content**.\n",
"\n",
"---\n",
"\n",
"## **🔹 Methods for Retrieval Ranking**\n",
"There are multiple ways to rank retrieved documents:\n",
"1. **Re-rank using an LLM** (best for complex queries).\n",
"2. **Use BM25 (Lexical Ranking)** alongside embeddings.\n",
"3. **Apply a trained re-ranking model** (e.g., `ColBERT`, `BGE-reranker`).\n",
"4. **Score documents using a similarity function**.\n",
"\n",
"---\n",
"\n",
"## **✅ Approach 1: LLM-Based Re-Ranking (Recommended)**\n",
"We can use an LLM to **re-rank** retrieved documents based on how well they match the query.\n",
"\n",
"### **🔹 Steps:**\n",
"1. Retrieve `k` documents from ChromaDB.\n",
"2. Ask the LLM to rank them by relevance.\n",
"3. Use the **top `n` ranked documents** in the final response generation.\n",
"\n",
"### **🔹 Implementation**\n",
"Modify your `query_index` function to **re-rank documents using an LLM**:\n",
"\n",
"```python\n",
"from langchain.prompts import PromptTemplate\n",
"from langchain.schema import AIMessage\n",
"from langchain.schema.runnable import RunnablePassthrough\n",
"import datetime\n",
"import logging\n",
"\n",
"logger = logging.getLogger(\"logger\")\n",
"\n",
"def query_index(vectordb, query, chain_type=\"stuff\", k=25, model_name=\"gemini-2.0-flash\", \n",
" date_filter=None, rerank_top_n=10):\n",
" \"\"\"\n",
" Queries the vectorstore with retrieval ranking.\n",
" \n",
" Args:\n",
" vectordb: The vector database.\n",
" query (str): The query string.\n",
" chain_type (str): The chain type.\n",
" k (int): The number of documents to retrieve.\n",
" model_name (str): The name of the language model.\n",
" date_filter (str, optional): A date string (YYYY-MM-DD) to filter documents by. Defaults to None.\n",
" rerank_top_n (int): Number of top documents to keep after re-ranking.\n",
" \n",
" Returns:\n",
" str: The answer from the language model.\n",
" \"\"\"\n",
" llm = ChatGoogleGenerativeAI(model=model_name, temperature=0.3) \n",
"\n",
" # Create the retriever with filter if date is provided\n",
" if date_filter:\n",
" try:\n",
" filter_date = datetime.datetime.strptime(date_filter, '%Y-%m-%d')\n",
" formatted_date = filter_date.strftime('%Y-%m-%d')\n",
" logger.info(f\"Date filter with query: {formatted_date}\")\n",
" except ValueError:\n",
" raise ValueError(\"Invalid date format. Please use YYYY-MM-DD.\")\n",
"\n",
" filter_criteria = {\"date\": { \"$eq\": formatted_date}}\n",
" retriever = vectordb.as_retriever(search_type=\"similarity\", search_kwargs={\"k\": k, \"filter\": filter_criteria})\n",
" else:\n",
" retriever = vectordb.as_retriever(search_kwargs={\"k\": k})\n",
"\n",
" # Retrieve initial `k` documents\n",
" retrieved_docs = retriever.invoke(query)\n",
" context = \"\\n\\n\".join([doc.page_content for doc in retrieved_docs]) \n",
" logger.info(f\"Retrieved {len(retrieved_docs)} documents for query: {query}\")\n",
"\n",
" # Define re-ranking prompt\n",
" rerank_prompt = PromptTemplate(\n",
" input_variables=[\"query\", \"documents\"],\n",
" template=(\n",
" \"You are an AI assistant ranking documents based on their relevance to the query.\\n\\n\"\n",
" \"Query: {query}\\n\\n\"\n",
" \"Documents:\\n{documents}\\n\\n\"\n",
" \"Rank the documents from most relevant to least relevant and return the top {rerank_top_n}.\"\n",
" ),\n",
" )\n",
"\n",
" # Rank retrieved documents using the LLM\n",
" rank_chain = rerank_prompt | llm | RunnablePassthrough()\n",
" ranked_response = rank_chain.invoke({\"query\": query, \"documents\": context, \"rerank_top_n\": rerank_top_n})\n",
"\n",
" # Extract ranked top `n` documents\n",
" ranked_docs = ranked_response.content.split(\"\\n\")[:rerank_top_n]\n",
" final_context = \"\\n\\n\".join(ranked_docs)\n",
"\n",
" # Define final QA prompt\n",
" qa_prompt = PromptTemplate(\n",
" input_variables=[\"context\", \"question\"],\n",
" template=(\n",
" \"You are an AI assistant retrieving factual and structured information.\\n\"\n",
" \"Use the following retrieved documents to answer the question accurately.\\n\\n\"\n",
" \"Context: {context}\\n\\n\"\n",
" \"Question: {question}\"\n",
" ),\n",
" )\n",
"\n",
" # Use RunnableSequence instead of deprecated LLMChain\n",
" final_chain = qa_prompt | llm | RunnablePassthrough()\n",
" response = final_chain.invoke({\"context\": final_context, \"question\": query})\n",
"\n",
" # Ensure response is extracted correctly\n",
" if isinstance(response, AIMessage):\n",
" return response.content \n",
" else:\n",
" return str(response) \n",
"```\n",
"\n",
"---\n",
"\n",
"### **🔹 What’s New in This Implementation?**\n",
"1. **Retrieve more documents (`k=25`) from ChromaDB.**\n",
"2. **Re-rank the documents using an LLM** to select the `rerank_top_n=10` most relevant.\n",
"3. **Use only top-ranked documents for the final response** to the user.\n",
"\n",
"---\n",
"\n",
"## **✅ Approach 2: Use BM25 for Hybrid Ranking**\n",
"If you want **lexical matching + embeddings**, use **BM25 + ChromaDB embeddings**:\n",
"\n",
"### **Steps**\n",
"1. Retrieve top `k` documents using **ChromaDB embeddings**.\n",
"2. Use **BM25 (bag-of-words ranking)** on these documents.\n",
"3. Take **top-ranked** documents and pass them to the LLM.\n",
"\n",
"### **Implementation**\n",
"- Install `rank_bm25`:\n",
" ```bash\n",
" pip install rank-bm25\n",
" ```\n",
"- Add BM25 re-ranking to your pipeline:\n",
" ```python\n",
" from rank_bm25 import BM25Okapi\n",
" from langchain.schema import Document\n",
"\n",
" def bm25_rerank(query, retrieved_docs, top_n=10):\n",
" \"\"\"Re-rank retrieved documents using BM25.\"\"\"\n",
" tokenized_docs = [doc.page_content.split() for doc in retrieved_docs]\n",
" bm25 = BM25Okapi(tokenized_docs)\n",
" scores = bm25.get_scores(query.split())\n",
"\n",
" # Sort documents based on BM25 score\n",
" ranked_docs = [retrieved_docs[i] for i in sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)]\n",
" return ranked_docs[:top_n]\n",
" ```\n",
"\n",
"Then, **modify `query_index` to use BM25 ranking** before passing docs to the LLM.\n",
"\n",
"---\n",
"\n",
"## **🔹 Which Approach Should You Use?**\n",
"| **Method** | **Pros** | **Cons** |\n",
"|------------------|---------|----------|\n",
"| **LLM-based Re-Ranking** (Approach 1) | Best for complex queries, uses language understanding | Slower, needs API calls |\n",
"| **BM25 + Embeddings (Approach 2)** | Improves lexical matching, fast & efficient | Doesn't handle semantics well |\n",
"| **ColBERT / BGE-reranker** | State-of-the-art accuracy, deep semantic ranking | Requires a fine-tuned model |\n",
"\n",
"For **best performance**, **combine both methods**: \n",
"1. Use **BM25 ranking first** to boost keyword relevance. \n",
"2. Then **use LLM-based ranking** on the BM25-ranked results. \n",
"\n",
"---\n",
"\n",
"## **✅ Final Takeaways**\n",
"- **Use LLM re-ranking** if you want **best semantic ranking**.\n",
"- **Use BM25 + embeddings** for **fast hybrid search**.\n",
"- **Try ColBERT/BGE-reranker** if you need **state-of-the-art retrieval ranking**.\n",
"\n",
"Would you like help setting up **ColBERT/BGE-reranker** in LangChain? 🚀"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "ml_env",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.11.9"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
|