Added fix to avoid irrelevant attributes being returned plus an example
of extracting unrelated entities and an exampe of using an 'extra_info'
attribute to extract unstructured data for an entity.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
"To extract entities, we need to create a schema like the following, were we specify all the properties we want to find and the type we expect them to have. We can also specify which of these properties are required and which are optional."
"To extract entities, we need to create a schema where we specify all the properties we want to find and the type we expect them to have. We can also specify which of these properties are required and which are optional."
]
]
},
},
{
{
"cell_type": "code",
"cell_type": "code",
"execution_count": 3,
"execution_count": 4,
"id": "4ac43eba",
"id": "4ac43eba",
"metadata": {},
"metadata": {},
"outputs": [],
"outputs": [],
"source": [
"schema = {\n",
" \"properties\": {\n",
" \"name\": {\"type\": \"string\"},\n",
" \"height\": {\"type\": \"integer\"},\n",
" \"hair_color\": {\"type\": \"string\"},\n",
" },\n",
" \"required\": [\"name\", \"height\"],\n",
"}"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "640bd005",
"metadata": {},
"outputs": [],
"source": [
"inp = \"\"\"\n",
"Alex is 5 feet tall. Claudia is 1 feet taller Alex and jumps higher than him. Claudia is a brunette and Alex is blonde.\n",
" \"\"\""
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "64313214",
"metadata": {},
"outputs": [],
"source": [
"chain = create_extraction_chain(schema, llm)"
]
},
{
"cell_type": "markdown",
"id": "17c48adb",
"metadata": {},
"source": [
"As we can see, we extracted the required entities and their properties in the required format (it even calculated Claudia's height before returning!)"
"Notice that we are using OpenAI functions under the hood and thus the model can only call one function per request (with one, unique schema)"
]
},
{
"cell_type": "markdown",
"id": "511b9838",
"metadata": {},
"source": [
"If we want to extract more than one entity type, we need to introduce a little hack - we will define our properties with an included entity type. \n",
"\n",
"Following we have an example where we also want to extract dog attributes from the passage. Notice the 'person_' and 'dog_' prefixes we use for each property; this tells the model which entity type the property refers to. In this way, the model can return properties from several entity types in one single call."
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "cf243a26",
"metadata": {},
"outputs": [],
"source": [
"source": [
"schema = {\n",
"schema = {\n",
" \"properties\": {\n",
" \"properties\": {\n",
@ -103,10 +198,10 @@
},
},
{
{
"cell_type": "markdown",
"cell_type": "markdown",
"id": "17c48adb",
"id": "eb074f7b",
"metadata": {},
"metadata": {},
"source": [
"source": [
"As we can see, we extracted the required entities and their properties in the required format:"
"People attributes and dog attributes were correctly extracted from the text in the same call"
]
]
},
},
{
{
@ -128,7 +223,207 @@
" 'person_hair_color': 'brunette'}]"
" 'person_hair_color': 'brunette'}]"
]
]
},
},
"execution_count": 6,
"execution_count": 11,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"chain.run(inp)"
]
},
{
"cell_type": "markdown",
"id": "0273e0e2",
"metadata": {},
"source": [
"## Unrelated entities"
]
},
{
"cell_type": "markdown",
"id": "c07b3480",
"metadata": {},
"source": [
"What if our entities are unrelated? In that case, the model will return the unrelated entities in different dictionaries, allowing us to successfully extract several unrelated entity types in the same call."
]
},
{
"cell_type": "markdown",
"id": "01d98af0",
"metadata": {},
"source": [
"Notice that we use `required: []`: we need to allow the model to return **only** person attributes or **only** dog attributes for a single entity (person or dog)"
"What if.. _we don't know what we want?_ More specifically, say we know a few properties we want to extract for a given entity but we also want to know if there's any extra information in the passage. Fortunately, we don't need to structure everything - we can have unstructured extraction as well. \n",
"\n",
"We can do this by introducing another hack, namely the *extra_info* attribute - let's see an example."