ContextualCompressionRetriever.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<img src=\"https://blog.langchain.dev/content/images/2023/04/Screenshot-2023-04-20-at-4.49.13-PM.png\" alt=\"A simple retrieval Q&A system\" width=\"1000\" height=\"400\" class=\"alignnone size-full wp-image-100\" >"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Created a chunk of size 1385, which is longer than the specified 300\n",
      "Created a chunk of size 491, which is longer than the specified 300\n",
      "Created a chunk of size 486, which is longer than the specified 300\n",
      "Created a chunk of size 480, which is longer than the specified 300\n",
      "Created a chunk of size 671, which is longer than the specified 300\n",
      "Created a chunk of size 1438, which is longer than the specified 300\n",
      "Created a chunk of size 506, which is longer than the specified 300\n",
      "Created a chunk of size 759, which is longer than the specified 300\n",
      "Created a chunk of size 615, which is longer than the specified 300\n",
      "Created a chunk of size 449, which is longer than the specified 300\n",
      "Created a chunk of size 471, which is longer than the specified 300\n",
      "Created a chunk of size 495, which is longer than the specified 300\n",
      "Created a chunk of size 715, which is longer than the specified 300\n",
      "Created a chunk of size 754, which is longer than the specified 300\n",
      "Created a chunk of size 752, which is longer than the specified 300\n",
      "Created a chunk of size 1078, which is longer than the specified 300\n",
      "Created a chunk of size 442, which is longer than the specified 300\n",
      "Created a chunk of size 802, which is longer than the specified 300\n",
      "Created a chunk of size 1510, which is longer than the specified 300\n",
      "Created a chunk of size 1607, which is longer than the specified 300\n",
      "Created a chunk of size 503, which is longer than the specified 300\n",
      "Created a chunk of size 973, which is longer than the specified 300\n",
      "Created a chunk of size 330, which is longer than the specified 300\n",
      "Created a chunk of size 632, which is longer than the specified 300\n",
      "Created a chunk of size 601, which is longer than the specified 300\n",
      "Created a chunk of size 504, which is longer than the specified 300\n",
      "Created a chunk of size 808, which is longer than the specified 300\n",
      "Created a chunk of size 848, which is longer than the specified 300\n",
      "Created a chunk of size 329, which is longer than the specified 300\n",
      "Created a chunk of size 319, which is longer than the specified 300\n",
      "Created a chunk of size 498, which is longer than the specified 300\n",
      "Created a chunk of size 395, which is longer than the specified 300\n",
      "Created a chunk of size 776, which is longer than the specified 300\n",
      "Created a chunk of size 373, which is longer than the specified 300\n",
      "Created a chunk of size 580, which is longer than the specified 300\n",
      "Created a chunk of size 384, which is longer than the specified 300\n",
      "Created a chunk of size 522, which is longer than the specified 300\n",
      "Created a chunk of size 362, which is longer than the specified 300\n",
      "Created a chunk of size 525, which is longer than the specified 300\n",
      "Created a chunk of size 584, which is longer than the specified 300\n",
      "Created a chunk of size 747, which is longer than the specified 300\n",
      "Created a chunk of size 496, which is longer than the specified 300\n",
      "Created a chunk of size 967, which is longer than the specified 300\n",
      "Created a chunk of size 507, which is longer than the specified 300\n",
      "Created a chunk of size 510, which is longer than the specified 300\n",
      "Created a chunk of size 789, which is longer than the specified 300\n",
      "Created a chunk of size 609, which is longer than the specified 300\n",
      "Created a chunk of size 673, which is longer than the specified 300\n",
      "Created a chunk of size 441, which is longer than the specified 300\n",
      "Created a chunk of size 515, which is longer than the specified 300\n",
      "Created a chunk of size 484, which is longer than the specified 300\n",
      "Created a chunk of size 574, which is longer than the specified 300\n",
      "Created a chunk of size 918, which is longer than the specified 300\n",
      "Created a chunk of size 757, which is longer than the specified 300\n",
      "Created a chunk of size 355, which is longer than the specified 300\n",
      "Created a chunk of size 520, which is longer than the specified 300\n",
      "Created a chunk of size 494, which is longer than the specified 300\n",
      "Created a chunk of size 448, which is longer than the specified 300\n",
      "Created a chunk of size 623, which is longer than the specified 300\n",
      "Created a chunk of size 963, which is longer than the specified 300\n",
      "Created a chunk of size 421, which is longer than the specified 300\n",
      "Created a chunk of size 318, which is longer than the specified 300\n",
      "Created a chunk of size 617, which is longer than the specified 300\n",
      "Created a chunk of size 358, which is longer than the specified 300\n",
      "Created a chunk of size 557, which is longer than the specified 300\n",
      "Created a chunk of size 935, which is longer than the specified 300\n",
      "Created a chunk of size 610, which is longer than the specified 300\n",
      "Created a chunk of size 624, which is longer than the specified 300\n",
      "Created a chunk of size 481, which is longer than the specified 300\n",
      "Created a chunk of size 834, which is longer than the specified 300\n",
      "Created a chunk of size 540, which is longer than the specified 300\n",
      "Created a chunk of size 591, which is longer than the specified 300\n",
      "Created a chunk of size 1024, which is longer than the specified 300\n",
      "Created a chunk of size 768, which is longer than the specified 300\n",
      "Created a chunk of size 483, which is longer than the specified 300\n",
      "Created a chunk of size 530, which is longer than the specified 300\n",
      "Created a chunk of size 457, which is longer than the specified 300\n",
      "Created a chunk of size 451, which is longer than the specified 300\n",
      "Created a chunk of size 357, which is longer than the specified 300\n",
      "Created a chunk of size 458, which is longer than the specified 300\n",
      "Created a chunk of size 520, which is longer than the specified 300\n",
      "Created a chunk of size 652, which is longer than the specified 300\n",
      "Created a chunk of size 384, which is longer than the specified 300\n",
      "Created a chunk of size 869, which is longer than the specified 300\n",
      "Created a chunk of size 535, which is longer than the specified 300\n",
      "Created a chunk of size 978, which is longer than the specified 300\n",
      "Created a chunk of size 592, which is longer than the specified 300\n",
      "Created a chunk of size 375, which is longer than the specified 300\n",
      "Created a chunk of size 638, which is longer than the specified 300\n",
      "Created a chunk of size 584, which is longer than the specified 300\n",
      "Created a chunk of size 388, which is longer than the specified 300\n",
      "Created a chunk of size 523, which is longer than the specified 300\n",
      "Created a chunk of size 408, which is longer than the specified 300\n",
      "Created a chunk of size 517, which is longer than the specified 300\n",
      "Created a chunk of size 838, which is longer than the specified 300\n",
      "Created a chunk of size 703, which is longer than the specified 300\n",
      "Created a chunk of size 1523, which is longer than the specified 300\n",
      "Created a chunk of size 1071, which is longer than the specified 300\n",
      "Created a chunk of size 588, which is longer than the specified 300\n",
      "Created a chunk of size 745, which is longer than the specified 300\n",
      "Created a chunk of size 589, which is longer than the specified 300\n",
      "Created a chunk of size 607, which is longer than the specified 300\n",
      "Created a chunk of size 321, which is longer than the specified 300\n",
      "Created a chunk of size 378, which is longer than the specified 300\n",
      "Created a chunk of size 464, which is longer than the specified 300\n",
      "Created a chunk of size 537, which is longer than the specified 300\n",
      "Created a chunk of size 402, which is longer than the specified 300\n",
      "Created a chunk of size 700, which is longer than the specified 300\n",
      "Created a chunk of size 364, which is longer than the specified 300\n",
      "Created a chunk of size 304, which is longer than the specified 300\n",
      "Created a chunk of size 424, which is longer than the specified 300\n",
      "Created a chunk of size 447, which is longer than the specified 300\n",
      "Created a chunk of size 425, which is longer than the specified 300\n",
      "Created a chunk of size 410, which is longer than the specified 300\n",
      "Created a chunk of size 512, which is longer than the specified 300\n",
      "Created a chunk of size 574, which is longer than the specified 300\n",
      "Created a chunk of size 457, which is longer than the specified 300\n",
      "Created a chunk of size 386, which is longer than the specified 300\n",
      "Created a chunk of size 821, which is longer than the specified 300\n",
      "Created a chunk of size 695, which is longer than the specified 300\n",
      "Created a chunk of size 424, which is longer than the specified 300\n",
      "Created a chunk of size 828, which is longer than the specified 300\n",
      "Created a chunk of size 311, which is longer than the specified 300\n",
      "Created a chunk of size 472, which is longer than the specified 300\n",
      "Created a chunk of size 473, which is longer than the specified 300\n",
      "Created a chunk of size 352, which is longer than the specified 300\n",
      "Created a chunk of size 562, which is longer than the specified 300\n",
      "Created a chunk of size 475, which is longer than the specified 300\n",
      "Created a chunk of size 508, which is longer than the specified 300\n",
      "Created a chunk of size 978, which is longer than the specified 300\n",
      "Created a chunk of size 835, which is longer than the specified 300\n",
      "Created a chunk of size 995, which is longer than the specified 300\n",
      "Created a chunk of size 412, which is longer than the specified 300\n",
      "Created a chunk of size 675, which is longer than the specified 300\n",
      "Created a chunk of size 777, which is longer than the specified 300\n",
      "Created a chunk of size 457, which is longer than the specified 300\n",
      "Created a chunk of size 565, which is longer than the specified 300\n",
      "Created a chunk of size 601, which is longer than the specified 300\n",
      "Created a chunk of size 518, which is longer than the specified 300\n",
      "Created a chunk of size 330, which is longer than the specified 300\n",
      "Created a chunk of size 371, which is longer than the specified 300\n",
      "Created a chunk of size 499, which is longer than the specified 300\n",
      "Created a chunk of size 349, which is longer than the specified 300\n",
      "Created a chunk of size 758, which is longer than the specified 300\n",
      "Created a chunk of size 303, which is longer than the specified 300\n",
      "Created a chunk of size 717, which is longer than the specified 300\n",
      "Created a chunk of size 823, which is longer than the specified 300\n",
      "Created a chunk of size 541, which is longer than the specified 300\n",
      "Created a chunk of size 477, which is longer than the specified 300\n",
      "Created a chunk of size 313, which is longer than the specified 300\n",
      "Created a chunk of size 479, which is longer than the specified 300\n",
      "Created a chunk of size 306, which is longer than the specified 300\n",
      "Created a chunk of size 624, which is longer than the specified 300\n",
      "Created a chunk of size 442, which is longer than the specified 300\n",
      "Created a chunk of size 789, which is longer than the specified 300\n",
      "Created a chunk of size 560, which is longer than the specified 300\n",
      "Created a chunk of size 511, which is longer than the specified 300\n",
      "Created a chunk of size 475, which is longer than the specified 300\n",
      "Created a chunk of size 402, which is longer than the specified 300\n",
      "Created a chunk of size 1060, which is longer than the specified 300\n",
      "Created a chunk of size 325, which is longer than the specified 300\n",
      "Created a chunk of size 1002, which is longer than the specified 300\n",
      "Created a chunk of size 1222, which is longer than the specified 300\n",
      "Created a chunk of size 560, which is longer than the specified 300\n",
      "Created a chunk of size 857, which is longer than the specified 300\n",
      "Created a chunk of size 841, which is longer than the specified 300\n",
      "Created a chunk of size 1173, which is longer than the specified 300\n",
      "Created a chunk of size 935, which is longer than the specified 300\n",
      "Created a chunk of size 474, which is longer than the specified 300\n",
      "Created a chunk of size 345, which is longer than the specified 300\n",
      "Created a chunk of size 382, which is longer than the specified 300\n",
      "Created a chunk of size 301, which is longer than the specified 300\n",
      "Created a chunk of size 482, which is longer than the specified 300\n",
      "Created a chunk of size 339, which is longer than the specified 300\n",
      "Created a chunk of size 474, which is longer than the specified 300\n",
      "Created a chunk of size 1417, which is longer than the specified 300\n",
      "Created a chunk of size 666, which is longer than the specified 300\n"
     ]
    }
   ],
   "source": [
    "from langchain.text_splitter import CharacterTextSplitter, RecursiveCharacterTextSplitter\n",
    "from langchain.embeddings import OpenAIEmbeddings, HuggingFaceEmbeddings\n",
    "from langchain.document_loaders import PDFMinerLoader, TextLoader\n",
    "from langchain.vectorstores import FAISS\n",
    "\n",
    "# Helper function for printing docs\n",
    "\n",
    "def pretty_print_docs(docs):\n",
    "    print(f\"\\n{'-' * 100}\\n\".join([f\"Document {i+1}:\\n\\n\" + d.page_content for i, d in enumerate(docs)]))\n",
    "\n",
    "documents = PDFMinerLoader(file_path=\"./arxiv_papers/2203.02155.pdf\").load()\n",
    "text_splitter = CharacterTextSplitter(chunk_size=300, chunk_overlap=0)\n",
    "texts = text_splitter.split_documents(documents=documents)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Document 1:\n",
      "\n",
      "Baselines. We compare the performance of our PPO models to our SFT models and GPT-3. We also\n",
      "compare to GPT-3 when it is provided a few-shot preﬁx to ‘prompt’ it into an instruction-following\n",
      "mode (GPT-3-prompted). This preﬁx is prepended to the user-speciﬁed instruction.6\n",
      "----------------------------------------------------------------------------------------------------\n",
      "Document 2:\n",
      "\n",
      "QA prompt\n",
      "instruction\n",
      "QA + instruct\n",
      "\n",
      "QA prompt\n",
      "instruction\n",
      "QA + instruct\n",
      "\n",
      "0.750\n",
      "0.774\n",
      "0.760\n",
      "\n",
      "0.448\n",
      "0.419\n",
      "0.420\n",
      "\n",
      "0.228\n",
      "0.211\n",
      "0.250\n",
      "\n",
      "0.312\n",
      "0.340\n",
      "0.335\n",
      "\n",
      "0.193\n",
      "0.212\n",
      "0.218\n",
      "\n",
      "0.721\n",
      "0.753\n",
      "0.773\n",
      "\n",
      "0.430\n",
      "0.413\n",
      "0.419\n",
      "\n",
      "0.229\n",
      "0.232\n",
      "0.261\n",
      "\n",
      "0.220\n",
      "0.414\n",
      "0.348\n",
      "\n",
      "0.186\n",
      "0.212\n",
      "0.267\n",
      "\n",
      "0.735\n",
      "0.796\n",
      "0.783\n",
      "\n",
      "0.410\n",
      "0.362\n",
      "0.353\n",
      "----------------------------------------------------------------------------------------------------\n",
      "Document 3:\n",
      "\n",
      "Figure 5: Comparing our models with FLAN and T0 in terms of Likert scores on a 1-7 scale, on the\n",
      "InstructGPT prompt distribution. FLAN and T0 perform better than default GPT-3, and comparably\n",
      "with a few-shot GPT-3 model placed into ‘instruction-following’ mode.\n",
      "----------------------------------------------------------------------------------------------------\n",
      "Document 4:\n",
      "\n",
      "InstructGPT models show promising generalization to instructions outside of the RLHF ﬁne-\n",
      "tuning distribution. We qualitatively probe InstructGPT’s capabilities, and ﬁnd that it is able to\n",
      "follow instructions for summarizing code, answer questions about code, and sometimes follows\n",
      "instructions in different languages, despite these instructions being very rare in the ﬁne-tuning\n",
      "distribution. In contrast, GPT-3 can perform these tasks but requires more careful prompting, and\n",
      "does not usually follow instructions in these domains. This result is exciting because it suggests that\n",
      "our models are able to generalize the notion of “following instructions.” They retain some alignment\n",
      "even on tasks for which they get very little direct supervision signal.\n"
     ]
    }
   ],
   "source": [
    "retriever = FAISS.from_documents(documents=texts, \n",
    "                                embedding=HuggingFaceEmbeddings()).as_retriever()\n",
    "\n",
    "docs = retriever.get_relevant_documents(\"What is Instruction Fine Tuning?\")\n",
    "pretty_print_docs(docs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.document_transformers import EmbeddingsRedundantFilter\n",
    "from langchain.retrievers.document_compressors import DocumentCompressorPipeline\n",
    "from langchain.text_splitter import CharacterTextSplitter\n",
    "from langchain.retrievers import ContextualCompressionRetriever\n",
    "\n",
    "\n",
    "splitter = CharacterTextSplitter(chunk_size=300, chunk_overlap=0, separator=\". \")\n",
    "redundant_filter = EmbeddingsRedundantFilter(embeddings=embeddings)\n",
    "relevant_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.76)\n",
    "pipeline_compressor = DocumentCompressorPipeline(\n",
    "    transformers=[splitter, redundant_filter, relevant_filter]\n",
    ")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "compression_retriever = ContextualCompressionRetriever(base_compressor=pipeline_compressor, base_retriever=retriever)\n",
    "compressed_docs = compression_retriever.get_relevant_documents(\"What did the president say about Ketanji Jackson Brown\")\n",
    "pretty_print_docs(compressed_docs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Document 1:\n",
      "\n",
      "Mixing in pretraining updates performs better than the simpler solution of increasing the KL co- efficient. In Figure 33, we show that there is a value of the pretraining mix coefficient that both reverses the performance regressions on SQuADv2 and DROP (the datasets we used for testing), and has minimal reductions in validation reward.\n",
      "----------------------------------------------------------------------------------------------------\n",
      "Document 2:\n",
      "\n",
      "Model samples on a prompt cherry-picked to show ability to follow instructions to describe code, along with random samples from the GPT-3 175B and InstructGPT 175B models.\n",
      "----------------------------------------------------------------------------------------------------\n",
      "Document 3:\n",
      "\n",
      "InstructGPT generalizes ‘following instructions’ to settings that we don’t supervise it in, for example on non-English language tasks and code-related tasks\n",
      "----------------------------------------------------------------------------------------------------\n",
      "Document 4:\n",
      "\n",
      "Labelers significantly prefer InstructGPT outputs over outputs from GPT-3. On our test set, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having over 100x fewer parameters. These models have the same architecture, and differ only by the fact that InstructGPT is fine-tuned on our human data.\n"
     ]
    }
   ],
   "source": [
    "#TODO: Add ContextualCompressionRetriever for compressing documents and querying them\n",
    "\n",
    "from langchain.llms import HuggingFacePipeline\n",
    "from langchain.retrievers import ContextualCompressionRetriever\n",
    "from langchain.retrievers.document_compressors import LLMChainExtractor\n",
    "from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM\n",
    "\n",
    "# local_llm = HuggingFacePipeline(pipeline=pipe)\n",
    "pipe = pipeline(\"text2text-generation\", model=\"google/flan-t5-large\", max_length=1000)\n",
    "llm = HuggingFacePipeline(pipeline=pipe)\n",
    "compressor = LLMChainExtractor.from_llm(llm)\n",
    "compression_retriever = ContextualCompressionRetriever(base_compressor=compressor, base_retriever=retriever)\n",
    "\n",
    "compressed_docs = compression_retriever.get_relevant_documents(\"What is Instruction Fine Tuning?\")\n",
    "pretty_print_docs(compressed_docs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "langchain",
   "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": "3.10.10"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}