Prompts

Beehive's prompts control how Beehive invokes chat models, interprets feedback, and routes conversations. These templates are rendered and stored in ~/.beehive/prompts/ when you create your first invokable. Prompt templates are Jinja2 templates. Variables are represented via the {{ ... }} expression. These variables get populated dynamically at runtime.

You can edit the wording of these prompts by directly modifying the text files this folder! Note that Beehive uses XML tags to structure prompts — we encourage you to maintain this style in your modifications.

Warning

If you modify the wording, do NOT add variables to the template. Additional variables will not be rendered properly, which may affect your invokable's behavior.

Here is a list of Beehive's internal prompts, separated

General

The following prompts are used across different kinds of invokables.

prompt	description
ask_question_prompt	used to explicitly allow invokables to ask questions to other invokables.
context_prompt_concise	used to represent context, i.e., messages from previous invokables. This prompt is "concise" — i.e., it excludes information about the format for the context.
context_prompt_full	used to represent context, i.e., messages from previous invokables. This prompt icludes information about the format for the context.
evaluation_prompt	used evaluate the quality of an invokable's output.
feedback_prompt	used to incorporate feedback to an invokable's task.
model_error_prompt	used when there is an error in the agent's output. For example, if the router's output does not meet a specific format, this prompt is used to inform the router of its mistake and to try again.

Beehive

The following prompts are used within the Beehive class.

prompt	description
router_prompting_prompt	used by a Beehive's router agent to create prompts for invokables in a Beehive.
router_question_prompt	used by a Beehive's router agent to direct a clarification question to invokable best posed to answer the question. This is used if a
router_routing_prompt	used by a Beehive's router agent to direct the conversation to the next appropriate invokable.

BeehiveDebateTeam

The following prompts are used within the BeehiveDebateTeam class.

prompt	description
debate_judge_prompt	used by the BeehiveDebateTeam's _judge_agent when judging the debater responses.
debate_judge_summary_prompt	used to represent the final answer from BeehiveDebateTeam. After the _judge_agent judges the debate rounds, Beehive uses this template to append the final answer to each of the member agents' states. This enables the debaters to build off of previous rounds.
long_form_debate_prompt	used to prompt members of a BeehiveDebateTeam to another round of debate.
short_form_debate_prompt	used to prompt members of a BeehiveDebateTeam to another round of debate.

BeehiveEnsemble

The following prompts are used within the BeehiveEnsemble class.

prompt	description
ensemble_func_summary_prompt	used to represent the final answer from BeehiveEnsemble when the ensemble uses a final_answer_method='similarity' to compute the final answer. Beehive uses this template to append the final answer to each of the member agents' states. This enables the debaters to build off of previous rounds.
ensemble_llm_summary_prompt	used to represent the final answer from BeehiveEnsemble when the ensemble uses a final_answer_method='llm' to create the final answer. Beehive uses this template to append the final answer to each of the member agents' states. This enables the debaters to build off of previous rounds.
synthesizer_prompt	used to synthesize the responses of ensemble members into a single, final answer.