Skip to content

Feature/improved-language-detection #525

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 18 commits into
base: main
Choose a base branch
from
Open

Feature/improved-language-detection #525

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
a350cb0
detect script as well
suzinyou Apr 2, 2025
5b17593
fix prompt
suzinyou Apr 2, 2025
f7fee04
fix tests
suzinyou Apr 9, 2025
ab2fd75
changes
suzinyou Apr 9, 2025
2cba258
remove search query during init chat history
suzinyou Apr 10, 2025
dbdeec4
fix tests and type
suzinyou Apr 10, 2025
09fdd6e
change schema and add validator
suzinyou Apr 10, 2025
ba013a5
fix test return mock value
suzinyou Apr 10, 2025
7ed1270
use enum not string
suzinyou Apr 10, 2025
e8cbf8a
fix type issues
suzinyou Apr 10, 2025
730a8cc
fix tests and always run paraphrase guardrail
suzinyou Apr 10, 2025
8b7e5a4
use uppercase
suzinyou Apr 10, 2025
15e083f
fix tests and how we get enum
suzinyou Apr 10, 2025
a4136be
add test cases
suzinyou Apr 10, 2025
e951d0a
clean up error logic
suzinyou Apr 10, 2025
c8117de
fix conftes
suzinyou Apr 10, 2025
a81a381
fix logic
suzinyou Apr 10, 2025
40187b8
Add query optimization back in, but don't run translation for chat qu…
suzinyou Apr 11, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
59 changes: 4 additions & 55 deletions .secrets.baseline
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -348,57 +348,6 @@
"line_number": 15
}
],
"core_backend/tests/api/conftest.py": [
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/conftest.py",
"hashed_secret": "407c6798fe20fd5d75de4a233c156cc0fce510e3",
"is_verified": false,
"line_number": 46
},
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/conftest.py",
"hashed_secret": "42553e798bc193bcf25368b5e53ec7cd771483a7",
"is_verified": false,
"line_number": 47
},
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/conftest.py",
"hashed_secret": "9fb7fe1217aed442b04c0f5e43b5d5a7d3287097",
"is_verified": false,
"line_number": 50
},
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/conftest.py",
"hashed_secret": "767ef7376d44bb6e52b390ddcd12c1cb1b3902a4",
"is_verified": false,
"line_number": 51
},
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/conftest.py",
"hashed_secret": "70240b5d0947cc97447de496284791c12b2e678a",
"is_verified": false,
"line_number": 56
},
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/conftest.py",
"hashed_secret": "80fea3e25cb7e28550d13af9dfda7a9bd08c1a78",
"is_verified": false,
"line_number": 57
},
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/conftest.py",
"hashed_secret": "3465834d516797458465ae4ed2c62e7020032c4e",
"is_verified": false,
"line_number": 317
}
],
"core_backend/tests/api/test.env": [
{
"type": "Secret Keyword",
Expand Down Expand Up @@ -448,14 +397,14 @@
"filename": "core_backend/tests/api/test_question_answer.py",
"hashed_secret": "1d2be5ef28a76e2207456e7eceabe1219305e43d",
"is_verified": false,
"line_number": 294
"line_number": 418
},
{
"type": "Secret Keyword",
"filename": "core_backend/tests/api/test_question_answer.py",
"hashed_secret": "6367c48dd193d56ea7b0baad25b19455e529f5ee",
"is_verified": false,
"line_number": 653
"line_number": 1018
}
],
"core_backend/tests/api/test_user_tools.py": [
Expand All @@ -473,7 +422,7 @@
"filename": "core_backend/tests/rails/test_language_identification.py",
"hashed_secret": "051b2c1d98174fabc4749641c4f4f4660556441e",
"is_verified": false,
"line_number": 48
"line_number": 69
}
],
"core_backend/tests/rails/test_paraphrasing.py": [
Expand Down Expand Up @@ -581,5 +530,5 @@
}
]
},
"generated_at": "2025-01-24T13:35:08Z"
"generated_at": "2025-04-10T13:44:48Z"
}
223 changes: 143 additions & 80 deletions core_backend/app/llm_call/llm_prompts.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,7 +103,7 @@
{context}

IMPORTANT NOTES ON THE "answer" FIELD:
- Answer in the language of the question ({original_language}).
- Answer in the language {original_language} in the script {original_script}.
- Answer should be concise, to the point, and no longer than 80 words.
- Do not include any information that is not present in the REFERENCE TEXT.
"""
Expand Down Expand Up @@ -182,6 +182,58 @@ class AlignmentScore(BaseModel):
model_config = ConfigDict(strict=True)


CHAT_RESPONSE_PROMPT = """\
You are an AI assistant designed to help users with their \
questions/concerns. You interact with users via a chat interface. You will \
be provided with ADDITIONAL RELEVANT INFORMATION that can address the \
user's questions/concerns.

BEFORE answering the user's LATEST MESSAGE, follow these steps:

1. Review the conversation history to ensure that you understand the \
context in which the user's LATEST MESSAGE is being asked.
2. Review the provided ADDITIONAL RELEVANT INFORMATION to ensure that you \
understand the most useful information related to the user's LATEST \
MESSAGE.

When you have completed the above steps, you will then write a JSON, whose \
TypeScript Interface is given below:

interface Response {{
extracted_info: string[];
answer: string;
}}

For "extracted_info", extract from the provided ADDITIONAL RELEVANT \
INFORMATION the most useful information related to the LATEST MESSAGE asked \
by the user, and list them one by one. If no useful information is found, \
return an empty list.

For "answer", understand the conversation history, ADDITIONAL RELEVANT \
INFORMATION, and the user's LATEST MESSAGE, and then provide an answer to \
the user's LATEST MESSAGE. If no useful information was found in the \
either the conversation history or the ADDITIONAL RELEVANT INFORMATION, \
respond with {failure_message}.

EXAMPLE RESPONSES:
{{"extracted_info": ["Pineapples are a blend of pinecones and apples.", \
"Pineapples have the shape of a pinecone."], \
"answer": "The 'pine-' from pineapples likely come from the fact that \
pineapples are a hybrid of pinecones and apples and its pinecone-like \
shape."}}
{{"extracted_info": [], "answer": "{failure_message}"}}

IMPORTANT NOTES ON THE "answer" FIELD:
- Keep in mind that the user is asking a {message_type} question.
- Answer in the language {original_language} in the script {original_script}.
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is the only line that actually changed. Evernthing else is exactly the same!

- Answer should be concise and to the point.
- Do not include any information that is not present in the ADDITIONAL \
RELEVANT INFORMATION.

Only output the JSON response, without any additional text.\
"""


class ChatHistory:
"""Contains the prompts and models for the chat history task."""

Expand Down Expand Up @@ -216,7 +268,7 @@ class ChatHistory:

{{
"message_type": "The type of the user's LATEST MESSAGE. List of valid
options are: {valid_message_types},
options are: {valid_message_types}",
"query": "The vector database query that you have constructed based on
the user's LATEST MESSAGE and the conversation history."
}}
Expand All @@ -227,62 +279,7 @@ class ChatHistory:
),
prompt_kws={"valid_message_types": _valid_message_types},
)
system_message_generate_response = format_prompt(
prompt=textwrap.dedent(
"""You are an AI assistant designed to help users with their
questions/concerns. You interact with users via a chat interface. You will
be provided with ADDITIONAL RELEVANT INFORMATION that can address the
user's questions/concerns.

BEFORE answering the user's LATEST MESSAGE, follow these steps:

1. Review the conversation history to ensure that you understand the
context in which the user's LATEST MESSAGE is being asked.
2. Review the provided ADDITIONAL RELEVANT INFORMATION to ensure that you
understand the most useful information related to the user's LATEST
MESSAGE.

When you have completed the above steps, you will then write a JSON, whose
TypeScript Interface is given below:

interface Response {{
extracted_info: string[];
answer: string;
}}

For "extracted_info", extract from the provided ADDITIONAL RELEVANT
INFORMATION the most useful information related to the LATEST MESSAGE asked
by the user, and list them one by one. If no useful information is found,
return an empty list.

For "answer", understand the conversation history, ADDITIONAL RELEVANT
INFORMATION, and the user's LATEST MESSAGE, and then provide an answer to
the user's LATEST MESSAGE. If no useful information was found in the
either the conversation history or the ADDITIONAL RELEVANT INFORMATION,
respond with {failure_message}.

EXAMPLE RESPONSES:
{{"extracted_info": [
"Pineapples are a blend of pinecones and apples.",
"Pineapples have the shape of a pinecone."
],
"answer": "The 'pine-' from pineapples likely come from the fact that
pineapples are a hybrid of pinecones and apples and its pinecone-like
shape."
}}
{{"extracted_info": [], "answer": "{failure_message}"}}

IMPORTANT NOTES ON THE "answer" FIELD:
- Keep in mind that the user is asking a {message_type} question.
- Answer in the language of the question ({original_language}).
- Answer should be concise and to the point.
- Do not include any information that is not present in the ADDITIONAL
RELEVANT INFORMATION.

Only output the JSON response, without any additional text.
"""
)
)
system_message_generate_response = CHAT_RESPONSE_PROMPT

class ChatHistoryConstructSearchQuery(BaseModel):
"""Pydantic model for the output of the construct search query chat history."""
Expand Down Expand Up @@ -337,22 +334,21 @@ class IdentifiedLanguage(str, Enum):
ENGLISH = "ENGLISH"
FRENCH = "FRENCH"
HINDI = "HINDI"
MARATHI = "MARATHI"
SWAHILI = "SWAHILI"
UNINTELLIGIBLE = "UNINTELLIGIBLE"
UNSUPPORTED = "UNSUPPORTED"

# XHOSA = "XHOSA"
# ZULU = "ZULU"

@classmethod
def get_supported_languages(cls) -> list[str]:
"""Return a list of supported languages.

Returns
-------
list[str]
A list of supported languages.
"""

return [
lang
for lang in cls._member_names_
Expand All @@ -377,31 +373,98 @@ def _missing_(cls, value: str) -> IdentifiedLanguage: # type: ignore[override]

return cls.UNSUPPORTED


class IdentifiedScript(str, Enum):
"""Script used in the user's input."""

LATIN = "LATIN"
DEVANAGARI = "DEVANAGARI"
BENGALI = "BENGALI"
TAMIL = "TAMIL"
TELUGU = "TELUGU"
KANNADA = "KANNADA"
MALAYALAM = "MALAYALAM"
GUJARATI = "GUJARATI"
# GURMUKHI = "GURMUKHI"
# ORIYA = "ORIYA"
# SINHALA = "SINHALA"
# MYANMAR = "MYANMAR"
# ETHIOPIC = "ETHIOPIC"
# GEORGIAN = "GEORGIAN"
# ARMENIAN = "ARMENIAN"
# HEBREW = "HEBREW"
# GREEK = "GREEK"
# TIBETAN = "TIBETAN"
# MONGOLIAN = "MONGOLIAN"
# KHMER = "KHMER"
# LAO = "LAO"
# VIETNAMESE = "VIETNAMESE"
# THAI_LAO = "THAI_LAO"
UNKNOWN = "UNKNOWN"

@classmethod
def get_prompt(cls) -> str:
"""Return the prompt for the language identification bot.
def get_supported_scripts(cls) -> list[str]:
"""Return a list of supported scripts.
Returns
-------
list[str]
A list of supported scripts.
"""
return [script for script in cls._member_names_ if script != "UNKNOWN"]

@classmethod
def _missing_(cls, value: str) -> IdentifiedScript: # type: ignore[override]
"""If script identified is not one of the supported scripts, it is
classified as UNKNOWN.
Parameters
----------
value
The script identified.
Returns
-------
str
The prompt for the language identification bot.
Script
The identified script (i.e., UNKNOWN).
"""
return cls.UNKNOWN

return textwrap.dedent(
f"""
You are a high-performing language identification bot that classifies the
language of the user input into one of {", ".join(cls._member_names_)}.

If the user input is
1. in one of the supported languages, then respond with that language.
2. written in a mix of languages, then respond with the dominant language.
3. in a real language but not a supported language, then respond with
UNSUPPORTED.
4. unintelligible or gibberish, then respond with UNINTELLIGIBLE.

Answer should be a single word and strictly one of
[{", ".join(cls._member_names_)}]"""
).strip()

class LanguageIdentificationResponse(BaseModel):
"""Pydantic model for the language identification response."""

language: IdentifiedLanguage
script: IdentifiedScript

model_config = ConfigDict(strict=True)


LANGUAGE_ID_PROMPT = (
f"""\
You are a high-performing language identification bot that classifies the \
language and script of the user input.

For each input, identify:
1. The language (must be one of {", ".join(IdentifiedLanguage._member_names_)})
2. The script (must be one of {", ".join(IdentifiedScript._member_names_)})

If the user input is:
1. in one of the supported languages, respond with that language and its script
2. written in a mix of languages, respond with the dominant language and its script
3. in a real language but not a supported language, respond with UNSUPPORTED and \
its script
4. unintelligible or gibberish, respond with UNINTELLIGIBLE and Latin"""
+ """
Examples:
"How many beds are there?" -> {{"language": "ENGLISH", "script": "LATIN"}}
"vahaan kitane bistar hain?" -> {{"language": "HINDI", "script": "LATIN"}}
"वहाँ कितने बिस्तर हैं?" -> {{"language": "HINDI", "script": "DEVANAGARI"}}
"Bonjour, comment allez-vous?" -> {{"language": "FRENCH", "script": "LATIN"}}
"Jambo, habari gani?" -> {{"language": "SWAHILI", "script": "LATIN"}}
"asdfjkl" -> {{"language": "UNINTELLIGIBLE", "script": "LATIN"}}
"مرحبا كيف حالك" -> {{"language": "UNSUPPORTED", "script": "ARABIC"}}

Respond with a JSON object containing "language" and "script" keys.
"""
)


class RAG(BaseModel):
Expand Down
Loading
Loading