updates to vectorstore memory (#2875)

docs/modules/memory/types/vectorstore_retriever_memory.ipynb

@@ -1,7 +1,6 @@
 {
  "cells": [
   {
-   "attachments": {},
    "cell_type": "markdown",
    "id": "ff4be5f3",
    "metadata": {},
@@ -10,7 +9,9 @@
     "\n",
     "`VectorStoreRetrieverMemory` stores memories in a VectorDB and queries the top-K most \"salient\" docs every time it is called.\n",
     "\n",
-    "This differs from most of the other Memory classes in that it doesn't explicitly track the order of interactions."
+    "This differs from most of the other Memory classes in that it doesn't explicitly track the order of interactions.\n",
+    "\n",
+    "In this case, the \"docs\" are previous conversation snippets. This can be useful to refer to relevant pieces of information that the AI was told earlier in the conversation."
    ]
   },
   {
@@ -26,7 +27,8 @@
     "from langchain.embeddings.openai import OpenAIEmbeddings\n",
     "from langchain.llms import OpenAI\n",
     "from langchain.memory import VectorStoreRetrieverMemory\n",
-    "from langchain.chains import ConversationChain"
+    "from langchain.chains import ConversationChain\n",
+    "from langchain.prompts import PromptTemplate"
    ]
   },
   {
@@ -41,7 +43,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": 29,
    "id": "eef56f65",
    "metadata": {
     "tags": []
@@ -61,7 +63,6 @@
    ]
   },
   {
-   "attachments": {},
    "cell_type": "markdown",
    "id": "8f4bdf92",
    "metadata": {},
@@ -73,7 +74,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 3,
+   "execution_count": 30,
    "id": "e00d4938",
    "metadata": {
     "tags": []
@@ -86,14 +87,14 @@
     "memory = VectorStoreRetrieverMemory(retriever=retriever)\n",
     "\n",
     "# When added to an agent, the memory object can save pertinent information from conversations or used tools\n",
-    "memory.save_context({\"input\": \"check the latest scores of the Warriors game\"}, {\"output\": \"the Warriors are up against the Astros 88 to 84\"})\n",
+    "memory.save_context({\"input\": \"My favorite food is pizza\"}, {\"output\": \"thats good to know\"})\n",
-    "memory.save_context({\"input\": \"I need help doing my taxes - what's the standard deduction this year?\"}, {\"output\": \"...\"})\n",
+    "memory.save_context({\"input\": \"My favorite sport is soccer\"}, {\"output\": \"...\"})\n",
-    "memory.save_context({\"input\": \"What's the the time?\"}, {\"output\": f\"It's {datetime.now()}\"}) # "
+    "memory.save_context({\"input\": \"I don't the Celtics\"}, {\"output\": \"ok\"}) # "
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 4,
+   "execution_count": 31,
    "id": "2fe28a28",
    "metadata": {
     "tags": []
@@ -103,7 +104,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "input: I need help doing my taxes - what's the standard deduction this year?\n",
+      "input: My favorite sport is soccer\n",
       "output: ...\n"
      ]
     }
@@ -111,7 +112,7 @@
    "source": [
     "# Notice the first result returned is the memory pertaining to tax help, which the language model deems more semantically relevant\n",
     "# to a 1099 than the other documents, despite them both containing numbers.\n",
-    "print(memory.load_memory_variables({\"prompt\": \"What's a 1099?\"})[\"history\"])"
+    "print(memory.load_memory_variables({\"prompt\": \"what sport should i watch?\"})[\"history\"])"
    ]
   },
   {
@@ -125,7 +126,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": 32,
    "id": "ebd68c10",
    "metadata": {
     "tags": []
@@ -141,9 +142,13 @@
       "Prompt after formatting:\n",
       "\u001b[32;1m\u001b[1;3mThe following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.\n",
       "\n",
+      "Relevant pieces of previous conversation:\n",
+      "input: My favorite food is pizza\n",
+      "output: thats good to know\n",
+      "\n",
+      "(You do not need to use these pieces of information if not relevant)\n",
+      "\n",
       "Current conversation:\n",
-      "input: I need help doing my taxes - what's the standard deduction this year?\n",
-      "output: ...\n",
       "Human: Hi, my name is Perry, what's up?\n",
       "AI:\u001b[0m\n",
       "\n",
@@ -153,18 +158,32 @@
     {
      "data": {
       "text/plain": [
-       "\" Hi Perry, my name is AI. I'm doing great, how about you? I understand you need help with your taxes. What specifically do you need help with?\""
+       "\" Hi Perry, I'm doing well. How about you?\""
       ]
      },
-     "execution_count": 5,
+     "execution_count": 32,
      "metadata": {},
      "output_type": "execute_result"
     }
    ],
    "source": [
     "llm = OpenAI(temperature=0) # Can be any valid LLM\n",
+    "_DEFAULT_TEMPLATE = \"\"\"The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.\n",
+    "\n",
+    "Relevant pieces of previous conversation:\n",
+    "{history}\n",
+    "\n",
+    "(You do not need to use these pieces of information if not relevant)\n",
+    "\n",
+    "Current conversation:\n",
+    "Human: {input}\n",
+    "AI:\"\"\"\n",
+    "PROMPT = PromptTemplate(\n",
+    "    input_variables=[\"history\", \"input\"], template=_DEFAULT_TEMPLATE\n",
+    ")\n",
     "conversation_with_summary = ConversationChain(\n",
     "    llm=llm, \n",
+    "    prompt=PROMPT,\n",
     "    # We set a very low max_token_limit for the purposes of testing.\n",
     "    memory=memory,\n",
     "    verbose=True\n",
@@ -174,7 +193,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 6,
+   "execution_count": 33,
    "id": "86207a61",
    "metadata": {
     "tags": []
@@ -190,10 +209,14 @@
       "Prompt after formatting:\n",
       "\u001b[32;1m\u001b[1;3mThe following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.\n",
       "\n",
+      "Relevant pieces of previous conversation:\n",
+      "input: My favorite sport is soccer\n",
+      "output: ...\n",
+      "\n",
+      "(You do not need to use these pieces of information if not relevant)\n",
+      "\n",
       "Current conversation:\n",
-      "input: check the latest scores of the Warriors game\n",
+      "Human: what's my favorite sport?\n",
-      "output: the Warriors are up against the Astros 88 to 84\n",
-      "Human: If the Cavaliers were to face off against the Warriers or the Astros, who would they most stand a chance to beat?\n",
       "AI:\u001b[0m\n",
       "\n",
       "\u001b[1m> Finished chain.\u001b[0m\n"
@@ -202,22 +225,22 @@
     {
      "data": {
       "text/plain": [
-       "\" It's hard to say without knowing the current form of the teams. However, based on the current scores, it looks like the Cavaliers would have a better chance of beating the Astros than the Warriors.\""
+       "' You told me earlier that your favorite sport is soccer.'"
       ]
      },
-     "execution_count": 6,
+     "execution_count": 33,
      "metadata": {},
      "output_type": "execute_result"
     }
    ],
    "source": [
     "# Here, the basketball related content is surfaced\n",
-    "conversation_with_summary.predict(input=\"If the Cavaliers were to face off against the Warriers or the Astros, who would they most stand a chance to beat?\")"
+    "conversation_with_summary.predict(input=\"what's my favorite sport?\")"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 7,
+   "execution_count": 34,
    "id": "8c669db1",
    "metadata": {
     "tags": []
@@ -233,10 +256,14 @@
       "Prompt after formatting:\n",
       "\u001b[32;1m\u001b[1;3mThe following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.\n",
       "\n",
+      "Relevant pieces of previous conversation:\n",
+      "input: My favorite food is pizza\n",
+      "output: thats good to know\n",
+      "\n",
+      "(You do not need to use these pieces of information if not relevant)\n",
+      "\n",
       "Current conversation:\n",
-      "input: What's the the time?\n",
+      "Human: Whats my favorite food\n",
-      "output: It's 2023-04-13 09:18:55.623736\n",
-      "Human: What day is it tomorrow?\n",
       "AI:\u001b[0m\n",
       "\n",
       "\u001b[1m> Finished chain.\u001b[0m\n"
@@ -245,10 +272,10 @@
     {
      "data": {
       "text/plain": [
-       "' Tomorrow is 2023-04-14.'"
+       "' You said your favorite food is pizza.'"
       ]
      },
-     "execution_count": 7,
+     "execution_count": 34,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -256,12 +283,12 @@
    "source": [
     "# Even though the language model is stateless, since relavent memory is fetched, it can \"reason\" about the time.\n",
     "# Timestamping memories and data is useful in general to let the agent determine temporal relevance\n",
-    "conversation_with_summary.predict(input=\"What day is it tomorrow?\")"
+    "conversation_with_summary.predict(input=\"Whats my favorite food\")"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 8,
+   "execution_count": 35,
    "id": "8c09a239",
    "metadata": {
     "tags": []
@@ -277,10 +304,14 @@
       "Prompt after formatting:\n",
       "\u001b[32;1m\u001b[1;3mThe following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.\n",
       "\n",
-      "Current conversation:\n",
+      "Relevant pieces of previous conversation:\n",
       "input: Hi, my name is Perry, what's up?\n",
-      "response:  Hi Perry, my name is AI. I'm doing great, how about you? I understand you need help with your taxes. What specifically do you need help with?\n",
+      "response:  Hi Perry, I'm doing well. How about you?\n",
-      "Human: What's your name?\n",
+      "\n",
+      "(You do not need to use these pieces of information if not relevant)\n",
+      "\n",
+      "Current conversation:\n",
+      "Human: What's my name?\n",
       "AI:\u001b[0m\n",
       "\n",
       "\u001b[1m> Finished chain.\u001b[0m\n"
@@ -289,10 +320,10 @@
     {
      "data": {
       "text/plain": [
-       "\" My name is AI. It's nice to meet you, Perry.\""
+       "' Your name is Perry.'"
       ]
      },
-     "execution_count": 8,
+     "execution_count": 35,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -301,8 +332,16 @@
     "# The memories from the conversation are automatically stored,\n",
     "# since this query best matches the introduction chat above,\n",
     "# the agent is able to 'remember' the user's name.\n",
-    "conversation_with_summary.predict(input=\"What's your name?\")"
+    "conversation_with_summary.predict(input=\"What's my name?\")"
    ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "df27c7dc",
+   "metadata": {},
+   "outputs": [],
+   "source": []
   }
  ],
  "metadata": {
@@ -321,7 +360,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.11.2"
+   "version": "3.9.1"
   }
  },
  "nbformat": 4,