Improve Llama v2 extract questions actor and associated prompt

- Format extract questions prompt format with newlines and whitespaces
- Make llama v2 extract questions prompt consistent

- Remove empty questions extracted by offline extract_questions actor
- Update implicit qs extraction unit test for offline search actor
This commit is contained in:
Debanjum Singh Solanky 2023-10-04 20:42:25 -07:00
parent a85ff941ca
commit 56bd69d5af
3 changed files with 26 additions and 20 deletions

View file

@ -113,7 +113,7 @@ def filter_questions(questions: List[str]):
]
filtered_questions = []
for q in questions:
if not any([word in q.lower() for word in hint_words]):
if not any([word in q.lower() for word in hint_words]) and not is_none_or_empty(q):
filtered_questions.append(q)
return filtered_questions

View file

@ -23,7 +23,7 @@ no_notes_found = PromptTemplate.from_template(
""".strip()
)
system_prompt_message_llamav2 = f"""You are Khoj, a friendly, smart and helpful personal assistant.
system_prompt_message_llamav2 = f"""You are Khoj, a smart, inquisitive and helpful personal assistant.
Using your general knowledge and our past conversations as context, answer the following question.
If you do not know the answer, say 'I don't know.'"""
@ -91,7 +91,7 @@ Question: {query}
notes_conversation_llamav2 = PromptTemplate.from_template(
"""
Notes:
User's Notes:
{references}
Question: {query}
""".strip()
@ -135,13 +135,19 @@ Answer (in second person):"""
extract_questions_llamav2_sample = PromptTemplate.from_template(
"""
<s>[INST] <<SYS>>Current Date: {current_date}<</SYS>> [/INST]</s>
<s>[INST]How was my trip to Cambodia?[/INST][]</s>
<s>[INST]Who did I visit the temple with on that trip?[/INST]Who did I visit the temple with in Cambodia?</s>
<s>[INST]How should I take care of my plants?[/INST]What kind of plants do I have? What issues do my plants have?</s>
<s>[INST]How many tennis balls fit in the back of a 2002 Honda Civic?[/INST]What is the size of a tennis ball? What is the trunk size of a 2002 Honda Civic?</s>
<s>[INST]What did I do for Christmas last year?[/INST]What did I do for Christmas {last_year} dt>='{last_christmas_date}' dt<'{next_christmas_date}'</s>
<s>[INST] How was my trip to Cambodia? [/INST]
How was my trip to Cambodia?</s>
<s>[INST] Who did I visit the temple with on that trip? [/INST]
Who did I visit the temple with in Cambodia?</s>
<s>[INST] How should I take care of my plants? [/INST]
What kind of plants do I have? What issues do my plants have?</s>
<s>[INST] How many tennis balls fit in the back of a 2002 Honda Civic? [/INST]
What is the size of a tennis ball? What is the trunk size of a 2002 Honda Civic?</s>
<s>[INST] What did I do for Christmas last year? [/INST]
What did I do for Christmas {last_year} dt>='{last_christmas_date}' dt<'{next_christmas_date}'</s>
<s>[INST] How are you feeling today? [/INST]</s>
<s>[INST]Is Alice older than Bob?[/INST]When was Alice born? What is Bob's age?</s>
<s>[INST] Is Alice older than Bob? [/INST]
When was Alice born? What is Bob's age?</s>
<s>[INST] <<SYS>>
Use these notes from the user's previous conversations to provide a response:
{chat_history}

View file

@ -128,15 +128,15 @@ def test_extract_multiple_explicit_questions_from_message(loaded_model):
@pytest.mark.chatquality
def test_extract_multiple_implicit_questions_from_message(loaded_model):
# Act
response = extract_questions_offline("Is Morpheus taller than Neo?", loaded_model=loaded_model)
response = extract_questions_offline("Is Carl taller than Ross?", loaded_model=loaded_model)
# Assert
expected_responses = ["height", "taller", "shorter", "heights"]
expected_responses = ["height", "taller", "shorter", "heights", "who"]
assert len(response) <= 3
for question in response:
assert any([expected_response in question.lower() for expected_response in expected_responses]), (
"Expected chat actor to ask follow-up questions about Morpheus and Neo, but got: " + question
"Expected chat actor to ask follow-up questions about Carl and Ross, but got: " + question
)