Split the retrieve_documents function in another cell for better expl…

…anation of the example

docs/docs/examples/qa-over-documents.ipynb

@@ -160,6 +160,53 @@
     ")"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Cool, now we can easily build a function that retrieves documents from our vector db:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 42,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "GPT4All LLMs\n",
+      "LLMs require a lot of GPU to run properly make it hard for the common folk to set one up locally. Fortunately, the folks at GPT4All are doing an excellent job in reall...\n",
+      "\n",
+      "LLMs\n",
+      "Large Language Models like GPT-4 is the whole reason LiteChain exists, we want to build on top of LLMs to construct an application. After learning the Chain Basics, it should ...\n",
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "from typing import AsyncGenerator\n",
+    "from litechain import as_async_generator\n",
+    "\n",
+    "def retrieve_documents(query: str, n_results: int) -> AsyncGenerator[str, None]:\n",
+    "    query_embedding = (\n",
+    "        openai.Embedding.create(model=\"text-embedding-ada-002\", input=query)\n",
+    "        .data[0]  # type: ignore\n",
+    "        .embedding\n",
+    "    )\n",
+    "\n",
+    "    results = collection.query(\n",
+    "        query_embeddings=query_embedding,\n",
+    "        n_results=n_results,\n",
+    "    )[\"documents\"]\n",
+    "    results = results[0]  # type: ignore\n",
+    "    return as_async_generator(*results)\n",
+    "\n",
+    "async for document in retrieve_documents(\"How to use GPT4All\", n_results=2):\n",
+    "    print(document[20:200] + \"...\\n\")"
+   ]
+  },
   {
    "cell_type": "markdown",
    "metadata": {},
@@ -171,7 +218,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "Finally, with all our documents indexed, we can now write a chain that will reply the user questions about those documents"
+    "Finally, with all our documents indexed and our query function, we can now write a chain that will reply the user questions about those documents using an LLM:"
    ]
   },
   {
@@ -211,26 +258,11 @@
     }
    ],
    "source": [
-    "from typing import AsyncGenerator, Iterable, List\n",
-    "from litechain import Chain, filter_final_output, as_async_generator\n",
+    "from typing import Iterable\n",
+    "from litechain import Chain, filter_final_output\n",
     "from litechain.contrib import OpenAIChatChain, OpenAIChatMessage, OpenAIChatDelta\n",
     "\n",
     "\n",
-    "def retrieve_documents(query: str, n_results: int) -> AsyncGenerator[str, None]:\n",
-    "    query_embedding = (\n",
-    "        openai.Embedding.create(model=\"text-embedding-ada-002\", input=query)\n",
-    "        .data[0]  # type: ignore\n",
-    "        .embedding\n",
-    "    )\n",
-    "\n",
-    "    results = collection.query(\n",
-    "        query_embeddings=query_embedding,\n",
-    "        n_results=n_results,\n",
-    "    )[\"documents\"]\n",
-    "    results = results[0]  # type: ignore\n",
-    "    return as_async_generator(*results)\n",
-    "\n",
-    "\n",
     "def chain(query):\n",
     "    return (\n",
     "        Chain[str, str](\n",
@@ -267,7 +299,7 @@
    "source": [
     "## Bonus: map-rerank document scoring\n",
     "\n",
-    "In the previous example, we used chromadb to return the top-3 results according to embeddings, however, sometimes you have a lot more documents, and the top-scoring provided by the embeddings on the vector-db might not be the best, at the same time, we cannot just send all the documents to the LLM due to context length limit.\n",
+    "In the previous example, we used chromadb to return the top-3 results according to embeddings, however, sometimes you have a lot more documents, and the top-scoring provided by the embeddings on the vector db might not be the best, at the same time, we cannot just send all the documents to the LLM due to context length limit.\n",
     "\n",
     "What we can do then is to leverage the LLM inteligence for better scoring, by returning more results from the vectordb search at first, and then ask the LLM to score each of them in parallel, to then pick just the best ones for the final answer.\n",
     "\n",