highperfocused/danswer
1
0
Fork 0
mirror of https://github.com/danswer-ai/danswer.git synced 2025-04-08 11:58:34 +02:00
Code Issues Packages Projects Releases Wiki Activity

minor touchups

Browse Source
This commit is contained in:
Yuhong Sun 2023-09-12 15:44:15 -07:00
parent 6f50f6710a
commit 9738e5e628
3 changed files with 14 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
backend/danswer/direct_qa/qa_block.py
View File

@ -110,13 +110,13 @@ class SingleMessageQAHandler(QAHandler):
content="You are a question answering system that is constantly learning and improving. "
"You can process and comprehend vast amounts of text and utilize this knowledge "
"to provide accurate and detailed answers to diverse queries.\n"
"You ALWAYS responds in a json containing an answer and quotes that support the answer.\n"
"You ALWAYS responds with only a json containing an answer and quotes that support the answer.\n"
"Your responses are as INFORMATIVE and DETAILED as possible.\n"
f"{GENERAL_SEP_PAT}CONTEXT:\n\n{context_docs_str}"
f"{GENERAL_SEP_PAT}Sample response:"
f"{CODE_BLOCK_PAT.format(json.dumps(EMPTY_SAMPLE_JSON))}\n"
f"{QUESTION_PAT} {query}\n"
"Hint: Make the answer as detailed as possible and use a JSON! "
"Hint: Make the answer as DETAILED as possible and respond in JSON format!\n"
"Quotes MUST be EXACT substrings from provided documents!"
)
]

11
backend/danswer/direct_qa/qa_utils.py
View File

@ -25,6 +25,7 @@ from danswer.dynamic_configs import get_dynamic_config_store
from danswer.llm.utils import check_number_of_tokens
from danswer.utils.logger import setup_logger
from danswer.utils.text_processing import clean_model_quote
from danswer.utils.text_processing import clean_up_code_blocks
from danswer.utils.text_processing import shared_precompare_cleanup
logger = setup_logger()
@ -158,7 +159,9 @@ def process_answer(
chunks: list[InferenceChunk],
is_json_prompt: bool = True,
) -> tuple[DanswerAnswer, DanswerQuotes]:
answer, quote_strings = separate_answer_quotes(answer_raw, is_json_prompt)
answer_clean = clean_up_code_blocks(answer_raw)
answer, quote_strings = separate_answer_quotes(answer_clean, is_json_prompt)
if answer == UNCERTAINTY_PAT or not answer:
if answer == UNCERTAINTY_PAT:
logger.debug("Answer matched UNCERTAINTY_PAT")
@ -226,6 +229,12 @@ def process_model_tokens(
# Then the chars after " will not be streamed, but this is ok as it prevents streaming the ? in the
# event that the model outputs the UNCERTAINTY_PAT
found_answer_start = True
# Prevent heavy cases of hallucinations where model is not even providing a json until later
if is_json_prompt and len(model_output) > 20:
logger.warning("LLM did not produce json as prompted")
found_answer_end = True
continue
if found_answer_start and not found_answer_end:

2
backend/danswer/secondary_llm_flows/query_validation.py
View File

@ -51,6 +51,8 @@ def get_query_validation_messages(user_query: str) -> list[dict[str, str]]:
f"relevant and consistent knowledge about the entity.\n"
f"The system is not tuned for writing code.\n"
f"The system is not tuned for interfacing with structured data via query languages like SQL.\n"
f"If the question might not require code or query language, "
f"then assume it can be answered without code or query language.\n"
f"Determine if that system should attempt to answer.\n"
f'"{ANSWERABLE_PAT}" must be exactly "True" or "False"\n{GENERAL_SEP_PAT}\n'
f"{ambiguous_example_question}{CODE_BLOCK_PAT.format(ambiguous_example_answer)}\n"