notice

This is unreleased documentation for Rasa Documentation Main/Unreleased version.
For the latest released documentation, see the latest version (3.x).

Version: Main/Unreleased

Knowledge Base Actions

Leverage information from knowledge bases inside conversations using `ActionQueryKnowledgeBase` in open source bot framework Rasa.

caution

This feature is experimental. We introduce experimental features to get feedback from our community, so we encourage you to try it out! However, the functionality might be changed or removed in the future. If you have feedback (positive or negative) please share it with us on the forum.

Knowledge base actions enable you to handle the following kind of conversations:

image

A common problem in conversational AI is that users do not only refer to certain objects by their names, but also use reference terms such as “the first one” or “it”. We need to keep track of the information that was presented to resolve these mentions to the correct object.

In addition, users may want to obtain detailed information about objects during a conversation – for example, whether a restaurant has outside seating, or how expensive it is. In order to respond to those user requests, knowledge about the restaurant domain is needed. Since the information is subject to change, hard-coding the information isn't the solution.

To handle the above challenges, Rasa can be integrated with knowledge bases. To use this integration, you can create a custom action that inherits from ActionQueryKnowledgeBase, a pre-written custom action that contains the logic to query a knowledge base for objects and their attributes.

You can find a complete example in examples/knowledgebasebot (knowledge base bot), as well as instructions for implementing this custom action below.

Using ActionQueryKnowledgeBase

Create a Knowledge Base

The data used to answer the user's requests will be stored in a knowledge base. A knowledge base can be used to store complex data structures. We suggest you get started by using the InMemoryKnowledgeBase. Once you want to start working with a large amount of data, you can switch to a custom knowledge base (see Creating Your Own Knowledge Base).

To initialize an InMemoryKnowledgeBase, you need to provide the data in a json file. The following example contains data about restaurants and hotels. The json structure should contain a key for every object type, i.e. "restaurant" and "hotel". Every object type maps to a list of objects – here we have a list of 3 restaurants and a list of 3 hotels.

{
"restaurant": [
{
"id": 0,
"name": "Donath",
"cuisine": "Italian",
"outside-seating": true,
"price-range": "mid-range"
},
{
"id": 1,
"name": "Berlin Burrito Company",
"cuisine": "Mexican",
"outside-seating": false,
"price-range": "cheap"
},
{
"id": 2,
"name": "I due forni",
"cuisine": "Italian",
"outside-seating": true,
"price-range": "mid-range"
}
],
"hotel": [
{
"id": 0,
"name": "Hilton",
"price-range": "expensive",
"breakfast-included": true,
"city": "Berlin",
"free-wifi": true,
"star-rating": 5,
"swimming-pool": true
},
{
"id": 1,
"name": "Hilton",
"price-range": "expensive",
"breakfast-included": true,
"city": "Frankfurt am Main",
"free-wifi": true,
"star-rating": 4,
"swimming-pool": false
},
{
"id": 2,
"name": "B&B",
"price-range": "mid-range",
"breakfast-included": false,
"city": "Berlin",
"free-wifi": false,
"star-rating": 1,
"swimming-pool": false
},
]
}

Once the data is defined in a json file, called, for example, data.json, you will be able use the this data file to create your InMemoryKnowledgeBase, which will be passed to the action that queries the knowledge base.

Every object in your knowledge base should have at least the "name" and "id" fields to use the default implementation. If it doesn't, you'll have to customize your InMemoryKnowledgeBase.

Define the NLU Data

In this section:

  • we will introduce a new intent, query_knowledge_base

  • we will annotate mention entities so that our model detects indirect mentions of objects like “the first one”

  • we will use synonyms extensively

For the bot to understand that the user wants to retrieve information from the knowledge base, you need to define a new intent. We will call it query_knowledge_base.

We can split requests that ActionQueryKnowledgeBase can handle into two categories: (1) the user wants to obtain a list of objects of a specific type, or (2) the user wants to know about a certain attribute of an object. The intent should contain lots of variations of both of these requests:

nlu:
- intent: query_knowledge_base
examples: |
- what [restaurants]{"entity": "object_type", "value": "restaurant"} can you recommend?
- list some [restaurants]{"entity": "object_type", "value": "restaurant"}
- can you name some [restaurants]{"entity": "object_type", "value": "restaurant"} please?
- can you show me some [restaurants]{"entity": "object_type", "value": "restaurant"} options
- list [German](cuisine) [restaurants]{"entity": "object_type", "value": "restaurant"}
- do you have any [mexican](cuisine) [restaurants]{"entity": "object_type", "value": "restaurant"}?
- do you know the [price range]{"entity": "attribute", "value": "price-range"} of [that one](mention)?
- what [cuisine](attribute) is [it](mention)?
- do you know what [cuisine](attribute) the [last one]{"entity": "mention", "value": "LAST"} has?
- does the [first one]{"entity": "mention", "value": "1"} have [outside seating]{"entity": "attribute", "value": "outside-seating"}?
- what is the [price range]{"entity": "attribute", "value": "price-range"} of [Berlin Burrito Company](restaurant)?
- what about [I due forni](restaurant)?
- can you tell me the [price range](attribute) of [that restaurant](mention)?
- what [cuisine](attribute) do [they](mention) have?

The above example just shows examples related to the restaurant domain. You should add examples for every object type that exists in your knowledge base to the same query_knowledge_base intent.

In addition to adding a variety of training examples for each query type, you need to specify and annotate the following entities in your training examples:

  • object_type: Whenever a training example references a specific object type from your knowledge base, the object type should be marked as an entity. Use synonyms to map e.g. restaurants to restaurant, the correct object type listed as a key in the knowledge base.

  • mention: If the user refers to an object via “the first one”, “that one”, or “it”, you should mark those terms as mention. We also use synonyms to map some of the mentions to symbols. You can learn about that in resolving mentions.

  • attribute: All attribute names defined in your knowledge base should be identified as attribute in the NLU data. Again, use synonyms to map variations of an attribute name to the one used in the knowledge base.

Remember to add those entities to your domain file (as entities and slots):

entities:
- object_type
- mention
- attribute
slots:
object_type:
type: any
influence_conversation: false
mappings:
- type: from_entity
entity: object_type
mention:
type: any
influence_conversation: false
mappings:
- type: from_entity
entity: mention
attribute:
type: any
influence_conversation: false
mappings:
- type: from_entity
entity: attribute

Create an Action to Query your Knowledge Base

To create your own knowledge base action, you need to inherit ActionQueryKnowledgeBase and pass the knowledge base to the constructor of ActionQueryKnowledgeBase.

from rasa_sdk.knowledge_base.storage import InMemoryKnowledgeBase
from rasa_sdk.knowledge_base.actions import ActionQueryKnowledgeBase
class MyKnowledgeBaseAction(ActionQueryKnowledgeBase):
def __init__(self):
knowledge_base = InMemoryKnowledgeBase("data.json")
super().__init__(knowledge_base)

Whenever you create an ActionQueryKnowledgeBase, you need to pass a KnowledgeBase to the constructor. It can be either an InMemoryKnowledgeBase or your own implementation of a KnowledgeBase (see Creating Your Own Knowledge Base). You can only pull information from one knowledge base, as the usage of multiple knowledge bases at the same time is not supported.

This is the entirety of the code for this action! The name of the action is action_query_knowledge_base. Don't forget to add it to your domain file:

actions:
- action_query_knowledge_base
note

If you overwrite the default action name action_query_knowledge_base, you need to add the following three unfeaturized slots to your domain file: knowledge_base_objects, knowledge_base_last_object, and knowledge_base_last_object_type. The slots are used internally by ActionQueryKnowledgeBase. If you keep the default action name, those slots will be automatically added for you.

You also need to make sure to add a story to your stories file that includes the intent query_knowledge_base and the action action_query_knowledge_base. For example:

stories:
- story: knowledge base happy path
steps:
- intent: greet
- action: utter_greet
- intent: query_knowledge_base
- action: action_query_knowledge_base
- intent: goodbye
- action: utter_goodbye

The last thing you need to do is to define the response utter_ask_rephrase in your domain file. If the action doesn't know how to handle the user's request, it will use this response to ask the user to rephrase. For example, add the following responses to your domain file:

responses:
utter_ask_rephrase:
- text: "Sorry, I'm not sure I understand. Could you rephrase it?"
- text: "Could you please rephrase your message? I didn't quite get that."

After adding all the relevant pieces, the action is now able to query the knowledge base.

How It Works

ActionQueryKnowledgeBase looks at both the entities that were picked up in the request as well as the previously set slots to decide what to query for.

Query the Knowledge Base for Objects

In order to query the knowledge base for any kind of object, the user's request needs to include the object type. Let's look at an example:

Can you please name some restaurants?

This question includes the object type of interest: “restaurant.” The bot needs to pick up on this entity in order to formulate a query – otherwise the action would not know what objects the user is interested in.

When the user says something like:

What Italian restaurant options in Berlin do I have?

The user wants to obtain a list of restaurants that (1) have Italian cuisine and (2) are located in Berlin. If the NER detects those attributes in the request of the user, the action will use those to filter the restaurants found in the knowledge base.

In order for the bot to detect these attributes, you need to mark “Italian” and “Berlin” as entities in the NLU data:

intents:
- intent: query_knowledge_base
examples: |
- What [Italian](cuisine) [restaurant](object_type) options in [Berlin](city) do I have?.

The names of the attributes, “cuisine” and “city,” should be equal to the ones used in the knowledge base. You also need to add those as entities and slots to the domain file.

Query the Knowledge Base for an Attribute of an Object

If the user wants to obtain specific information about an object, the request should include both the object and attribute of interest.

New in 3.6

The user is not required to query the knowledge base to list any kind of object prior to this. The ActionQueryKnowledgeBase will extract the object type from the user's request and query the knowledge base for an attribute of the object.

For example, if the user asks something like:

What is the cuisine of Berlin Burrito Company?

The user wants to obtain the “cuisine” (attribute of interest) for the restaurant “Berlin Burrito Company” (object of interest).

The attribute and object of interest should be marked as entities in the NLU training data:

intents:
- intent: query_knowledge_base
examples: |
- What is the [cuisine](attribute) of [Berlin Burrito Company](restaurant)?

Make sure to add the object type, “restaurant,” to the domain file as entity and slot. This will support ActionQueryKnowledgeBase to extract the object type of the object the user is interested in.

Resolve Mentions

Following along from the above example, users may not always refer to restaurants by their names. Users can either refer to the object of interest by its name, e.g. “Berlin Burrito Company” (representation string of the object), or they may refer to a previously listed object via a mention, for example:

What is the cuisine of the second restaurant you mentioned?

Our action is able to resolve these mentions to the actual object in the knowledge base. More specifically, it can resolve two mention types: (1) ordinal mentions, such as “the first one”, and (2) mentions such as “it” or “that one”.

Ordinal Mentions

When a user refers to an object by its position in a list, it is called an ordinal mention. Here's an example:

  • User: What restaurants in Berlin do you know?

  • Bot: Found the following objects of type 'restaurant': 1: I due forni 2: PastaBar 3: Berlin Burrito Company

  • User: Does the first one have outside seating?

The user referred to “I due forni” by the term “the first one”. Other ordinal mentions might include “the second one,” “the last one,” “any,” or “3”.

Ordinal mentions are typically used when a list of objects was presented to the user. To resolve those mentions to the actual object, we use an ordinal mention mapping which is set in the KnowledgeBase class. The default mapping looks like:

{
"1": lambda l: l[0],
"2": lambda l: l[1],
"3": lambda l: l[2],
"4": lambda l: l[3],
"5": lambda l: l[4],
"6": lambda l: l[5],
"7": lambda l: l[6],
"8": lambda l: l[7],
"9": lambda l: l[8],
"10": lambda l: l[9],
"ANY": lambda l: random.choice(l),
"LAST": lambda l: l[-1],
}

The ordinal mention mapping maps a string, such as “1”, to the object in a list, e.g. lambda l: l[0], meaning the object at index 0.

As the ordinal mention mapping does not, for example, include an entry for “the first one”, it is important that you use Entity Synonyms to map “the first one” in your NLU data to “1”:

intents:
- intent: query_knowledge_base
examples: |
- Does the [first one]{entity: "mention", value": 1} have [outside seating]{entity: "attribute", value": "outside-seating"}

The NER detects “first one” as a mention entity, but puts “1” into the mention slot. Thus, our action can take the mention slot together with the ordinal mention mapping to resolve “first one” to the actual object “I due forni”.

You can overwrite the ordinal mention mapping by calling the function set_ordinal_mention_mapping() on your KnowledgeBase implementation (see Customizing the InMemoryKnowledgeBase).

Other Mentions

Take a look at the following conversation:

  • User: What is the cuisine of PastaBar?

  • Bot: PastaBar has an Italian cuisine.

  • User: Does it have wifi?

  • Bot: Yes.

  • User: Can you give me an address?

In the question “Does it have wifi?”, the user refers to “PastaBar” by the word “it”. If the NER detected “it” as the entity mention, the knowledge base action would resolve it to the last mentioned object in the conversation, “PastaBar”.

In the next input, the user refers indirectly to the object “PastaBar” instead of mentioning it explicitly. The knowledge base action would detect that the user wants to obtain the value of a specific attribute, in this case, the address. If no mention or object was detected by the NER, the action assumes the user is referring to the most recently mentioned object, “PastaBar”.

You can disable this behavior by setting use_last_object_mention to False when initializing the action.

Customization

Customizing ActionQueryKnowledgeBase

You can overwrite the following two functions of ActionQueryKnowledgeBase if you'd like to customize what the bot says to the user:

  • utter_objects()

  • utter_attribute_value()

utter_objects() is used when the user has requested a list of objects. Once the bot has retrieved the objects from the knowledge base, it will respond to the user by default with a message, formatted like:

Found the following objects of type 'restaurant': 1: I due forni 2: PastaBar 3: Berlin Burrito Company

Or, if no objects are found,

I could not find any objects of type 'restaurant'.

If you want to change the utterance format, you can overwrite the method utter_objects() in your action.

The function utter_attribute_value() determines what the bot utters when the user is asking for specific information about an object.

If the attribute of interest was found in the knowledge base, the bot will respond with the following utterance:

'Berlin Burrito Company' has the value 'Mexican' for attribute 'cuisine'.

If no value for the requested attribute was found, the bot will respond with

Did not find a valid value for attribute 'cuisine' for object 'Berlin Burrito Company'.

If you want to change the bot utterance, you can overwrite the method utter_attribute_value().

note

There is a tutorial on our blog about how to use knowledge bases in custom actions. The tutorial explains the implementation behind ActionQueryKnowledgeBase in detail.

Creating Your Own Knowledge Base Actions

ActionQueryKnowledgeBase should allow you to easily get started with integrating knowledge bases into your actions. However, the action can only handle two kind of user requests:

  • the user wants to get a list of objects from the knowledge base

  • the user wants to get the value of an attribute for a specific object

The action is not able to compare objects or consider relations between objects in your knowledge base. Furthermore, resolving any mention to the last mentioned object in the conversation might not always be optimal.

If you want to tackle more complex use cases, you can write your own custom action. We added some helper functions to rasa_sdk.knowledge_base.utils (link to code ) to help you when implement your own solution. We recommend using KnowledgeBase interface so that you can still use the ActionQueryKnowledgeBase alongside your new custom action.

If you write a knowledge base action that tackles one of the above use cases or a new one, be sure to tell us about it on the forum!

Customizing the InMemoryKnowledgeBase

The class InMemoryKnowledgeBase inherits KnowledgeBase. You can customize your InMemoryKnowledgeBase by overwriting the following functions:

  • get_key_attribute_of_object(): To keep track of what object the user was talking about last, we store the value of the key attribute in a specific slot. Every object should have a key attribute that is unique, similar to the primary key in a relational database. By default, the name of the key attribute for every object type is set to id. You can overwrite the name of the key attribute for a specific object type by calling set_key_attribute_of_object().

  • get_representation_function_of_object(): Let's focus on the following restaurant:

    {
    "id": 0,
    "name": "Donath",
    "cuisine": "Italian",
    "outside-seating": true,
    "price-range": "mid-range"
    }

    When the user asks the bot to list any Italian restaurant, it doesn't need all of the details of the restaurant. Instead, you want to provide a meaningful name that identifies the restaurant – in most cases, the name of the object will do. The function get_representation_function_of_object() returns a lambda function that maps the above restaurant object to its name.

    lambda obj: obj["name"]

    This function is used whenever the bot is talking about a specific object, so that the user is presented a meaningful name for the object.

    By default, the lambda function returns the value of the "name" attribute of the object. If your object does not have a "name" attribute , or the "name" of an object is ambiguous, you should set a new lambda function for that object type by calling set_representation_function_of_object().

  • set_ordinal_mention_mapping(): The ordinal mention mapping is needed to resolve an ordinal mention, such as “second one,” to an object in a list. By default, the ordinal mention mapping looks like this:

    {
    "1": lambda l: l[0],
    "2": lambda l: l[1],
    "3": lambda l: l[2],
    "4": lambda l: l[3],
    "5": lambda l: l[4],
    "6": lambda l: l[5],
    "7": lambda l: l[6],
    "8": lambda l: l[7],
    "9": lambda l: l[8],
    "10": lambda l: l[9],
    "ANY": lambda l: random.choice(l),
    "LAST": lambda l: l[-1],
    }

    You can overwrite it by calling the function set_ordinal_mention_mapping(). If you want to learn more about how this mapping is used, check out Resolve Mentions.

See the example bot for an example implementation of an InMemoryKnowledgeBase that uses the method set_representation_function_of_object() to overwrite the default representation of the object type “hotel.” The implementation of the InMemoryKnowledgeBase itself can be found in the rasa-sdk package.

Creating Your Own Knowledge Base

If you have more data or if you want to use a more complex data structure that, for example, involves relations between different objects, you can create your own knowledge base implementation. Just inherit KnowledgeBase and implement the methods get_objects(), get_object(), get_object_types() and get_attributes_of_object(). The knowledge base code provides more information on what those methods should do.

You can also customize your knowledge base further, by adapting the methods mentioned in the section Customizing the InMemoryKnowledgeBase.

note

We wrote a blog post that explains how you can set up your own knowledge base.