highperfocused/danswer
1
0
Fork 0
mirror of https://github.com/danswer-ai/danswer.git synced 2025-03-17 21:32:36 +01:00
Code Issues Packages Projects Releases Wiki Activity

Yuhong

Browse Source
This commit is contained in:
Yuhong Sun 2025-02-02 16:14:18 -08:00 committed by Evan Lohn
parent a067b32467
commit 506a9f1b94
6 changed files with 149 additions and 125 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
backend/onyx/agents/agent_search/deep_search/main/nodes/extract_entities_terms.py
View File

@ -23,9 +23,8 @@ from onyx.agents.agent_search.shared_graph_utils.utils import (
get_langgraph_node_log_string,
)
from onyx.configs.constants import NUM_EXPLORATORY_DOCS
from onyx.prompts.agent_search import (
ENTITY_TERM_EXTRACTION_PROMPT,
)
from onyx.prompts.agent_search import ENTITY_TERM_EXTRACTION_PROMPT
from onyx.prompts.agent_search import ENTITY_TERM_EXTRACTION_PROMPT_JSON_EXAMPLE
def extract_entities_terms(
@ -58,16 +57,21 @@ def extract_entities_terms(
# start with the entity/term/extraction
doc_context = format_docs(initial_search_docs)
# Calculation here is only approximate
doc_context = trim_prompt_piece(
graph_config.tooling.fast_llm.config,
doc_context,
ENTITY_TERM_EXTRACTION_PROMPT + question,
ENTITY_TERM_EXTRACTION_PROMPT
+ question
+ ENTITY_TERM_EXTRACTION_PROMPT_JSON_EXAMPLE,
)
msg = [
HumanMessage(
content=ENTITY_TERM_EXTRACTION_PROMPT.format(
question=question, context=doc_context
),
)
+ ENTITY_TERM_EXTRACTION_PROMPT_JSON_EXAMPLE,
)
]
fast_llm = graph_config.tooling.fast_llm

39
backend/onyx/agents/agent_search/shared_graph_utils/agent_prompt_ops.py
View File

@ -13,6 +13,7 @@ from onyx.agents.agent_search.shared_graph_utils.utils import (
from onyx.agents.agent_search.shared_graph_utils.utils import remove_document_citations
from onyx.agents.agent_search.shared_graph_utils.utils import summarize_history
from onyx.configs.agent_configs import AGENT_MAX_STATIC_HISTORY_WORD_LENGTH
from onyx.configs.constants import MessageType
from onyx.context.search.models import InferenceSection
from onyx.llm.interfaces import LLMConfig
from onyx.llm.utils import get_max_input_tokens
@ -21,6 +22,9 @@ from onyx.natural_language_processing.utils import tokenizer_trim_content
from onyx.prompts.agent_search import HISTORY_FRAMING_PROMPT
from onyx.prompts.agent_search import SUB_QUESTION_RAG_PROMPT
from onyx.prompts.prompt_utils import build_date_time_string
from onyx.utils.logger import setup_logger
logger = setup_logger()
def build_sub_question_answer_prompt(
@ -36,9 +40,9 @@ def build_sub_question_answer_prompt(
date_str = build_date_time_string()
# TODO: This should include document metadata and title
docs_format_list = [
f"""Document Number: [D{doc_num + 1}]\n
Content: {doc.combined_content}\n\n"""
f"Document Number: [D{doc_num + 1}]\nContent: {doc.combined_content}\n\n"
for doc_num, doc in enumerate(docs)
]
@ -88,7 +92,6 @@ def trim_prompt_piece(config: LLMConfig, prompt_piece: str, reserved_str: str) -
def build_history_prompt(config: GraphConfig, question: str) -> str:
prompt_builder = config.inputs.prompt_builder
model = config.tooling.fast_llm
persona_base = get_persona_agent_prompt_expressions(
config.inputs.search_request.persona
).base_prompt
@ -102,23 +105,31 @@ def build_history_prompt(config: GraphConfig, question: str) -> str:
history_components = []
previous_message_type = None
for message in prompt_builder.raw_message_history:
if "user" in message.message_type:
if message.message_type == MessageType.USER:
history_components.append(f"User: {message.message}\n")
previous_message_type = "user"
elif "assistant" in message.message_type:
# only use the last agent answer for the history
if previous_message_type != "assistant":
history_components.append(f"You/Agent: {message.message}\n")
else:
history_components = history_components[:-1]
history_components.append(f"You/Agent: {message.message}\n")
previous_message_type = "assistant"
previous_message_type = MessageType.USER
elif message.message_type == MessageType.ASSISTANT:
# Previously there could be multiple assistant messages in a row
# Now this is handled at the message history construction
assert previous_message_type is not MessageType.ASSISTANT
history_components.append(f"You/Agent: {message.message}\n")
previous_message_type = MessageType.ASSISTANT
else:
# Other message types are not included here, currently there should be no other message types
logger.error(
f"Unhandled message type: {message.message_type} with message: {message.message}"
)
continue
history = "\n".join(history_components)
history = remove_document_citations(history)
if len(history.split()) > AGENT_MAX_STATIC_HISTORY_WORD_LENGTH:
history = summarize_history(history, question, persona_base, model)
history = summarize_history(
history=history,
question=question,
persona_specification=persona_base,
llm=config.tooling.fast_llm,
)
return HISTORY_FRAMING_PROMPT.format(history=history) if history else ""

2
backend/onyx/agents/agent_search/shared_graph_utils/models.py
View File

@ -119,7 +119,7 @@ class CombinedAgentMetrics(BaseModel):
class PersonaPromptExpressions(BaseModel):
contextualized_prompt: str
base_prompt: str
base_prompt: str | None
class AgentPromptEnrichmentComponents(BaseModel):

31
backend/onyx/agents/agent_search/shared_graph_utils/utils.py
View File

@ -56,6 +56,7 @@ from onyx.prompts.agent_search import (
from onyx.prompts.agent_search import (
HISTORY_CONTEXT_SUMMARY_PROMPT,
)
from onyx.prompts.prompt_utils import handle_onyx_date_awareness
from onyx.tools.force import ForceUseTool
from onyx.tools.tool_constructor import SearchToolConfig
from onyx.tools.tool_implementations.search.search_tool import (
@ -227,16 +228,26 @@ def get_test_config(
def get_persona_agent_prompt_expressions(
persona: Persona | None,
) -> PersonaPromptExpressions:
if persona is None:
persona_base = ""
persona_prompt = ASSISTANT_SYSTEM_PROMPT_DEFAULT
else:
persona_base = "\n".join([x.system_prompt for x in persona.prompts])
persona_prompt = ASSISTANT_SYSTEM_PROMPT_PERSONA.format(
persona_prompt=persona_base
if persona is None or len(persona.prompts) == 0:
# TODO base_prompt should be None, but no time to properly fix
return PersonaPromptExpressions(
contextualized_prompt=ASSISTANT_SYSTEM_PROMPT_DEFAULT, base_prompt=""
)
# Only a 1:1 mapping between personas and prompts currently
prompt = persona.prompts[0]
prompt_config = PromptConfig.from_model(prompt)
datetime_aware_system_prompt = handle_onyx_date_awareness(
prompt_str=prompt_config.system_prompt,
prompt_config=prompt_config,
add_additional_info_if_no_tag=prompt.datetime_aware,
)
return PersonaPromptExpressions(
contextualized_prompt=persona_prompt, base_prompt=persona_base
contextualized_prompt=ASSISTANT_SYSTEM_PROMPT_PERSONA.format(
persona_prompt=datetime_aware_system_prompt
),
base_prompt=datetime_aware_system_prompt,
)
@ -322,7 +333,7 @@ def get_answer_citation_ids(answer_str: str) -> list[int]:
def summarize_history(
history: str, question: str, persona_specification: str, model: LLM
history: str, question: str, persona_specification: str | None, llm: LLM
) -> str:
history_context_prompt = remove_document_citations(
HISTORY_CONTEXT_SUMMARY_PROMPT.format(
@ -332,7 +343,7 @@ def summarize_history(
)
)
history_response = model.invoke(history_context_prompt)
history_response = llm.invoke(history_context_prompt)
assert isinstance(history_response.content, str)
return history_response.content

13
backend/onyx/chat/chat_utils.py
View File

@ -166,6 +166,7 @@ def create_chat_chain(
)
current_message: ChatMessage | None = root_message
previous_message: ChatMessage | None = None
while current_message is not None:
child_msg = current_message.latest_child_message
@ -183,7 +184,17 @@ def create_chat_chain(
"could not find next message in the same session"
)
mainline_messages.append(current_message)
if (
current_message.message_type == MessageType.ASSISTANT
and previous_message is not None
and previous_message.message_type == MessageType.ASSISTANT
and mainline_messages
):
mainline_messages[-1] = current_message
else:
mainline_messages.append(current_message)
previous_message = current_message
if not mainline_messages:
raise RuntimeError("Could not trace chat message history")

175
backend/onyx/prompts/agent_search.py
View File

@ -2,7 +2,6 @@
SEPARATOR_LINE = "-------"
UNKNOWN_ANSWER = "I do not have enough information to answer this question."
NO_RECOVERED_DOCS = "No relevant information recovered"
DATE_PROMPT = "Today is {date}.\n\n"
SUB_CHECK_YES = "yes"
SUB_CHECK_NO = "no"
@ -16,9 +15,7 @@ For more context, here is the history of the conversation so far that preceded t
""".strip()
ASSISTANT_SYSTEM_PROMPT_DEFAULT = (
"""You are an assistant for question-answering tasks."""
)
ASSISTANT_SYSTEM_PROMPT_DEFAULT = "You are an assistant for question-answering tasks."
ASSISTANT_SYSTEM_PROMPT_PERSONA = f"""
You are an assistant for question-answering tasks. Here is more information about you:
@ -28,21 +25,25 @@ You are an assistant for question-answering tasks. Here is more information abou
""".strip()
SUB_QUESTION_ANSWER_TEMPLATE = """\n
Sub-Question: Q{sub_question_num}\n Sub-Question:\n - \n{sub_question}\n --\nAnswer:\n -\n {sub_answer}\n\n
"""
SUB_QUESTION_ANSWER_TEMPLATE = f"""
Sub-Question: Q{{sub_question_num}}
Question:
{{sub_question}}
{SEPARATOR_LINE}
Answer:
{{sub_answer}}
""".strip()
SUB_QUESTION_ANSWER_TEMPLATE_REFINED = f"""
Sub-Question: Q{{sub_question_num}}\n
Type:
{SEPARATOR_LINE}
{{sub_question_type}}
{SEPARATOR_LINE}
Sub-Question: Q{{sub_question_num}}
Type: {{sub_question_type}}
Sub-Question:
{SEPARATOR_LINE}
{{sub_question}}
{SEPARATOR_LINE}
Answer:
{SEPARATOR_LINE}
{{sub_answer}}
@ -73,30 +74,33 @@ And here is the context retrieved:
{SEPARATOR_LINE}
Please format your answer as a json object in the following format:
{{
"retrieved_entities_relationships": {{
""".strip()
ENTITY_TERM_EXTRACTION_PROMPT_JSON_EXAMPLE = """
{
"retrieved_entities_relationships": {
"entities": [
{{
{
"entity_name": "<assign a name for the entity>",
"entity_type": "<specify a short type name for the entity, such as 'company', 'location',...>"
}}
}
],
"relationships": [
{{
{
"relationship_name": "<assign a name for the relationship>",
"relationship_type": "<specify a short type name for the relationship, such as 'sales_to', 'is_location_of',...>",
"relationship_entities": ["<related entity name 1>", "<related entity name 2>", "..."]
}}
}
],
"terms": [
{{
{
"term_name": "<assign a name for the term>",
"term_type": "<specify a short type name for the term, such as 'revenue', 'market_share',...>",
"term_similar_to": ["<list terms that are similar to this term>"]
}}
}
]
}}
}}
}
}
""".strip()
@ -259,79 +263,63 @@ SUB_QUESTION_RAG_PROMPT = (
SUB_ANSWER_CHECK_PROMPT = (
"""\n
Your task is to see whether a given answer addresses a given question.
Please do not use any internal knowledge you may have - just focus on whether the answer
as given seems to largely address the question as given, or at least addresses part of the question.
Here is the question:
\n-------\n
{question}
\n-------\n
Here is the suggested answer:
\n-------\n
{base_answer}
\n-------\n
Does the suggested answer address the question? Please answer with """
+ f'"{SUB_CHECK_YES}" or "{SUB_CHECK_NO}".'
)
"Determine whether the given answer addresses the given question. "
"Please do not use any internal knowledge you may have - just focus on whether the answer "
"as given seems to largely address the question as given, or at least addresses part of the question.\n\n"
"Here is the question:\n"
f"{SEPARATOR_LINE}\n"
"{question}\n"
f"{SEPARATOR_LINE}\n\n"
"Here is the suggested answer:\n"
f"{SEPARATOR_LINE}\n"
"{base_answer}\n"
f"{SEPARATOR_LINE}\n\n"
f'Does the suggested answer address the question? Please answer with "{SUB_CHECK_YES}" or "{SUB_CHECK_NO}".'
).strip()
# Initial Answer Generation
INITIAL_ANSWER_PROMPT_W_SUB_QUESTIONS = (
""" \n
{persona_specification}
{date_prompt}
Use the information provided below - and only the provided information - to answer the provided main question.
The information provided below consists of:
1) a number of answered sub-questions - these are very important to help you organize your thoughts and your answer
2) a number of documents that deemed relevant for the question.
{history}
It is critical that you provide prover inline citations to documents in the format [[D1]](), [[D2]](), [[D3]](), etc.!
It is important that the citation is close to the information it supports. If you have multiple citations that support
a fact, please cite for example as [[D1]]()[[D3]](), or [[D2]]()[[D4]](), etc.
Feel free to also cite sub-questions in addition to documents, but make sure that you have documents cited with the sub-question
citation. If you want to cite both a document and a sub-question, please use [[D1]]()[[Q3]](), or [[D2]]()[[D7]]()[[Q4]](), etc.
Again, please NEVER cite sub-questions without a document citation!
Proper citations are very important for the user!
IMPORTANT RULES:
- If you cannot reliably answer the question solely using the provided information, say that you cannot reliably answer.
You may give some additional facts you learned, but do not try to invent an answer.
- If the information is empty or irrelevant, just say """
+ f'"{UNKNOWN_ANSWER}"'
+ """.
- If the information is relevant but not fully conclusive, specify that the information is not conclusive and say why.
Again, you should be sure that the answer is supported by the information provided!
Try to keep your answer concise. But also highlight uncertainties you may have should there be substantial ones,
or assumptions you made.
Here is the contextual information:
---------------
*Answered Sub-questions (these should really matter!):
\n-------\n
{answered_sub_questions}
\n-------\n
And here are relevant document information that support the sub-question answers, or that are relevant for the actual question:\n
\n-------\n
{relevant_docs}
\n-------\n
And here is the question I want you to answer based on the information above:
\n-------\n
{question}
\n-------\n\n
Please keep your answer brief and concise, and focus on facts and data.
Answer:"""
)
"{persona_specification}\n\n"
"Use the information provided below - and only the provided information - to answer the provided main question.\n\n"
"The information provided below consists of:\n"
" 1) a number of answered sub-questions - these are very important to help you organize your thoughts and your answer\n"
" 2) a number of documents that deemed relevant for the question.\n\n"
"{history}\n\n"
"It is critical that you provide prover inline citations to documents in the format [[D1]](), [[D2]](), [[D3]](), etc.!\n"
"It is important that the citation is close to the information it supports. If you have multiple citations that support\n"
"a fact, please cite for example as [[D1]]()[[D3]](), or [[D2]]()[[D4]](), etc.\n"
"Feel free to also cite sub-questions in addition to documents, but make sure that you have documents cited with the "
"sub-question citation. If you want to cite both a document and a sub-question, please use [[D1]]()[[Q3]](), or "
"[[D2]]()[[D7]]()[[Q4]](), etc.\n"
"Again, please NEVER cite sub-questions without a document citation! "
"Proper citations are very important for the user!\n\n"
"IMPORTANT RULES:\n"
" - If you cannot reliably answer the question solely using the provided information, say that you cannot reliably answer.\n"
" You may give some additional facts you learned, but do not try to invent an answer.\n"
f' - If the information is empty or irrelevant, just say "{UNKNOWN_ANSWER}".\n'
" - If the information is relevant but not fully conclusive, specify that the information is not conclusive and say why.\n\n"
"Again, you should be sure that the answer is supported by the information provided!\n\n"
"Try to keep your answer concise. But also highlight uncertainties you may have should there be substantial ones,\n"
"or assumptions you made.\n\n"
"Here is the contextual information:\n"
"---------------\n\n"
"*Answered Sub-questions (these should really matter!):\n"
f"{SEPARATOR_LINE}\n"
"{answered_sub_questions}\n"
f"{SEPARATOR_LINE}\n\n"
"And here are relevant document information that support the sub-question answers, "
"or that are relevant for the actual question:\n"
f"{SEPARATOR_LINE}\n"
"{relevant_docs}\n"
f"{SEPARATOR_LINE}\n\n"
"And here is the question I want you to answer based on the information above:\n"
f"{SEPARATOR_LINE}\n"
"{question}\n"
f"{SEPARATOR_LINE}\n\n"
"Please keep your answer brief and concise, and focus on facts and data.\n\n"
"Answer:"
).strip()
# used if sub_question_answer_str is empty
@ -339,7 +327,6 @@ INITIAL_ANSWER_PROMPT_WO_SUB_QUESTIONS = (
"""\n
{answered_sub_questions}
{persona_specification}
{date_prompt}
Use the information provided below - and only the provided information - to answer the provided question.
The information provided below consists of a number of documents that were deemed relevant for the question.
@ -465,7 +452,7 @@ Generate the list of questions separated by one new line like this:
REFINED_ANSWER_PROMPT_W_SUB_QUESTIONS = (
"""\n
{persona_specification}
{date_prompt}
Your task is to improve on a given answer to a question, as the initial answer was found to be lacking in some way.
Use the information provided below - and only the provided information - to write your new and improved answer.
@ -542,7 +529,7 @@ REFINED_ANSWER_PROMPT_WO_SUB_QUESTIONS = (
"""\n
{answered_sub_questions}\n
{persona_specification}
{date_prompt}
Use the information provided below - and only the
provided information - to answer the provided question.