Prompting language models to select API calls
A communications session with a user may be automated using a language model. The language model may be instructed to select a next action to be performed where the next action may include transmitting a responsive communication to the user or performing an API call. The prompt used to query the language model may include one or more of the following: a representation of text of the communications session, a list of available API calls, instructions to select a next action, a representation of API calls performed, or a representation of API call responses received. The language model may be sequentially queried to continue the communications session by transmitting responsive communications or performing API calls. In some implementations, a prompt template may be used to generate the prompt and a prompt template may be selected using text of the communications session.
1. A computer-implemented method, comprising:
starting a communications session with a user;
obtaining session text of the communications session;
creating a first language model prompt, wherein the first language model prompt comprises: (a) a representation of the session text, (b) a list of available API calls, and (c) instructions to select a next action from a plurality of available next actions, wherein the plurality of available next actions comprises (i) transmitting a responsive communication and (ii) performing an API call;
submitting the first language model prompt to a language model to obtain a first language model response comprising a first next action;
determining that the first next action corresponds to performing a first API call;
performing the first API call to obtain a first API call response;
creating a second language model prompt, wherein the second language model prompt comprises: (a) the representation of the session text, (b) the list of available API calls, (c) the instructions to select a next action from the plurality of available next actions, and (d) a text representation of the first API call response;
submitting the second language model prompt to the language model to obtain a second language model response comprising a second next action;
determining that the second next action corresponds to transmitting a first responsive communication; and
transmitting the first responsive communication to the user.
2. The computer-implemented method of claim 1 , wherein obtaining session text comprises performing automatic speech recognition of an audio communication.
3. The computer-implemented method of claim 1 , wherein the second language model prompt comprises a text representation of the first API call.
4. The computer-implemented method of claim 1 , comprising:
selecting the list of available API calls using the session text.
5. The computer-implemented method of claim 4 , wherein selecting the list of available API calls comprises processing at least a portion of the session text with a classifier neural network.
6. The computer-implemented method of claim 4 , wherein selecting the list of available API calls comprises submitting a third language model prompt to the language model or a second language model.
7. The computer-implemented method of claim 1 , wherein:
the plurality of available next actions comprises selecting a language model prompt template;
selecting the language model prompt template; and
creating a third language model prompt using the language model prompt template.
8. The computer-implemented method of claim 1 , wherein the second language model prompt comprises an expected format of an API call response.
9. A system, comprising at least one server computer comprising at least one processor and at least one memory, the at least one server computer configured to:
start a communications session with a user;
obtain session text of the communications session;
create a first language model prompt, wherein the first language model prompt comprises: (a) a representation of the session text, (b) a list of available API calls, and (c) instructions to select a next action from a plurality of available next actions, wherein the plurality of available next actions comprises (i) transmitting a responsive communication and (ii) performing an API call;
submit the first language model prompt to a language model to obtain a first language model response comprising a first next action;
determine that the first next action corresponds to performing a first API call;
perform the first API call to obtain a first API call response;
create a second language model prompt, wherein the second language model prompt comprises: (a) the representation of the session text, (b) the list of available API calls, (c) the instructions to select a next action from the plurality of available next actions, and (d) a text representation of the first API call response;
submit the second language model prompt to the language model to obtain a second language model response comprising a second next action;
determine that the second next action corresponds to transmitting a first responsive communication; and
transmit the first responsive communication to the user.
10. The system of claim 9 , wherein the language model is provided by a third party and submitting the first language model prompt to the language model comprises transmitting the first language model prompt to the third party.
11. The system of claim 9 , the at least one server computer is configured to:
select a language model prompt template using the session text; and
create the first language model prompt using the language model prompt template.
12. The system of claim 11 , wherein selecting the language model prompt template comprises processing at least a portion of the session text with a classifier neural network.
13. The system of claim 11 , wherein the at least one server computer is configured to select the language model prompt template by submitting a third language model prompt to the language model or a second language model.
14. The system of claim 11 , wherein:
the plurality of available next actions comprises selecting a second list of available API calls; and
the at least one server computer is configured to:
select the second list of available API calls, and
create a third language model prompt using the second list of available API calls.
15. The system of claim 9 , wherein the list of available API calls comprises function names and corresponding function arguments.
16. The system of claim 15 , wherein the list of available API calls comprises function argument types.
17. One or more non-transitory, computer-readable media comprising computer-executable instructions that, when executed, cause at least one processor to perform actions comprising:
starting a communications session with a user;
obtaining session text of the communications session;
creating a first language model prompt, wherein the first language model prompt comprises: (a) a representation of the session text, (b) a list of available API calls, and (c) instructions to select a next action from a plurality of available next actions, wherein the plurality of available next actions comprises (i) transmitting a responsive communication and (ii) performing an API call;
submitting the first language model prompt to a language model to obtain a first language model response comprising a first next action;
determining that the first next action corresponds to performing a first API call;
performing the first API call to obtain a first API call response;
creating a second language model prompt, wherein the second language model prompt comprises: (a) the representation of the session text, (b) the list of available API calls, (c) the instructions to select a next action from the plurality of available next actions, and (d) a text representation of the first API call response;
submitting the second language model prompt to the language model to obtain a second language model response comprising a second next action;
determining that the second next action corresponds to transmitting a first responsive communication; and
transmitting the first responsive communication to the user.
18. The one or more non-transitory, computer-readable media of claim 17 , wherein the user is obtaining customer support from a company.
19. The one or more non-transitory, computer-readable media of claim 17 , wherein the first language model prompt does not include a graph corresponding to an interactive voice response system.
20. The one or more non-transitory, computer-readable media of claim 17 , wherein the representation of the session text comprises an automatically-generated summary of at least a portion of the session text.