Generative Characters (langchain-ai#2859)

Add a time-weighted memory retriever and a notebook that approximates a Generative Agent from https://arxiv.org/pdf/2304.03442.pdf The "daily plan" components are removed for now since they are less useful without a virtual world, but the memory is an interesting component to build off. --------- Co-authored-by: Harrison Chase <[email protected]>
baohang · Apr 17, 2023 · 99c0382 · 99c0382
1 parent a9310a3
commit 99c0382
Show file tree

Hide file tree

Showing 7 changed files with 1,771 additions and 3 deletions.
diff --git a/docs/modules/indexes/retrievers/examples/time_weighted_vectorstore.ipynb b/docs/modules/indexes/retrievers/examples/time_weighted_vectorstore.ipynb
@@ -0,0 +1,213 @@
+{
+ "cells": [
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "id": "a90b7557",
+   "metadata": {},
+   "source": [
+    "# Time Weighted VectorStore Retriever\n",
+    "\n",
+    "This retriever uses a combination of semantic similarity and recency.\n",
+    "\n",
+    "The algorithm for scoring them is:\n",
+    "\n",
+    "```\n",
+    "semantic_similarity + (1.0 - decay_rate) ** hours_passed\n",
+    "```\n",
+    "\n",
+    "Notably, hours_passed refers to the hours passed since the object in the retriever **was last accessed**, not since it was created. This means that frequently accessed objects remain \"fresh.\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "f22cc96b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import faiss\n",
+    "\n",
+    "from datetime import datetime, timedelta\n",
+    "from langchain.docstore import InMemoryDocstore\n",
+    "from langchain.embeddings import OpenAIEmbeddings\n",
+    "from langchain.retrievers import TimeWeightedVectorStoreRetriever\n",
+    "from langchain.schema import Document\n",
+    "from langchain.vectorstores import FAISS\n"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "id": "6af7ea6b",
+   "metadata": {},
+   "source": [
+    "## Low Decay Rate\n",
+    "\n",
+    "A low decay rate (in this, to be extreme, we will set close to 0) means memories will be \"remembered\" for longer. A decay rate of 0 means memories never be forgotten, making this retriever equivalent to the vector lookup."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "c10e7696",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Define your embedding model\n",
+    "embeddings_model = OpenAIEmbeddings()\n",
+    "# Initialize the vectorstore as empty\n",
+    "embedding_size = 1536\n",
+    "index = faiss.IndexFlatL2(embedding_size)\n",
+    "vectorstore = FAISS(embeddings_model.embed_query, index, InMemoryDocstore({}), {})\n",
+    "retriever = TimeWeightedVectorStoreRetriever(vectorstore=vectorstore, decay_rate=.0000000000000000000000001, k=1) "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "86dbadb9",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "['129ba56b-7e7f-480b-83b3-8138a7f5db4a']"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "yesterday = datetime.now() - timedelta(days=1)\n",
+    "retriever.add_documents([Document(page_content=\"hello world\", metadata={\"last_accessed_at\": yesterday})])\n",
+    "retriever.add_documents([Document(page_content=\"hello foo\")])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "a580be32",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[Document(page_content='hello foo', metadata={'last_accessed_at': datetime.datetime(2023, 4, 16, 15, 46, 43, 860748), 'created_at': datetime.datetime(2023, 4, 16, 15, 46, 14, 469670), 'buffer_idx': 1})]"
+      ]
+     },
+     "execution_count": 9,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# \"Hello World\" is returned first because it is most salient, and the decay rate is close to 0., meaning it's still recent enough\n",
+    "retriever.get_relevant_documents(\"hello world\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "id": "ca056896",
+   "metadata": {},
+   "source": [
+    "## High Decay Rate\n",
+    "\n",
+    "With a high decay factor (e.g., several 9's), the recency score quickly goes to 0! If you set this all the way to 1, recency is 0 for all objects, once again making this equivalent to a vector lookup.\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "dc37669b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Define your embedding model\n",
+    "embeddings_model = OpenAIEmbeddings()\n",
+    "# Initialize the vectorstore as empty\n",
+    "embedding_size = 1536\n",
+    "index = faiss.IndexFlatL2(embedding_size)\n",
+    "vectorstore = FAISS(embeddings_model.embed_query, index, InMemoryDocstore({}), {})\n",
+    "retriever = TimeWeightedVectorStoreRetriever(vectorstore=vectorstore, decay_rate=.999, k=1) "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "fa284384",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "['8fff7ef8-3a30-40f3-b42e-b8d5c7850863']"
+      ]
+     },
+     "execution_count": 7,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "yesterday = datetime.now() - timedelta(days=1)\n",
+    "retriever.add_documents([Document(page_content=\"hello world\", metadata={\"last_accessed_at\": yesterday})])\n",
+    "retriever.add_documents([Document(page_content=\"hello foo\")])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "7558f94d",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[Document(page_content='hello foo', metadata={'last_accessed_at': datetime.datetime(2023, 4, 16, 15, 46, 17, 646927), 'created_at': datetime.datetime(2023, 4, 16, 15, 46, 14, 469670), 'buffer_idx': 1})]"
+      ]
+     },
+     "execution_count": 8,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# \"Hello Foo\" is returned first because \"hello world\" is mostly forgotten\n",
+    "retriever.get_relevant_documents(\"hello world\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "bf6d8c90",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "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": "3.11.2"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}