pieter/eveAI
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

- Improvements to enable deployment in the cloud, mainly changing file access to Minio

Browse Source 
- Improvements on RAG logging, and some debugging in that area
This commit is contained in:
Josako
2024-08-01 17:35:54 +02:00
parent 88ca04136d
commit 64cf8df3a9
19 changed files with 617 additions and 206 deletions
Download Patch File Download Diff File Expand all files Collapse all files

90
config/config.py
View File

@@ -61,7 +61,7 @@ class Config(object):
# supported LLMs
SUPPORTED_EMBEDDINGS = ['openai.text-embedding-3-small', 'openai.text-embedding-3-large', 'mistral.mistral-embed']
SUPPORTED_LLMS = ['openai.gpt-4o', 'anthropic.claude-3-5-sonnet']
SUPPORTED_LLMS = ['openai.gpt-4o', 'anthropic.claude-3-5-sonnet', 'openai.gpt-4o-mini']
ANTHROPIC_LLM_VERSIONS = {'claude-3-5-sonnet': 'claude-3-5-sonnet-20240620', }
@@ -71,6 +71,7 @@ class Config(object):
# Annotation text chunk length
ANNOTATION_TEXT_CHUNK_LENGTH = {
'openai.gpt-4o': 10000,
'openai.gpt-4o-mini': 10000,
'anthropic.claude-3-5-sonnet': 8000
}
@@ -184,12 +185,95 @@ class DevConfig(Config):
# PATH settings
ffmpeg_path = '/usr/bin/ffmpeg'
# MINIO
MINIO_ENDPOINT = 'minio:9000'
MINIO_ACCESS_KEY = 'minioadmin'
MINIO_SECRET_KEY = 'minioadmin'
class ProdConfig(Config):
DEVELOPMENT = False
DEBUG = False
# SQLALCHEMY_DATABASE_URI = environ.get('SQLALCHEMY_DATABASE_URI') or \
# 'sqlite:///' + os.path.join(basedir, 'db.sqlite')
DEVELOPMENT = True
DEBUG = True
FLASK_DEBUG = True
PYCHARM_DEBUG = False
DB_HOST = environ.get('DB_HOST', 'bswnz4.stackhero-network.com')
DB_USER = environ.get('DB_USER', 'luke_skywalker')
DB_PASS = environ.get('DB_PASS', '2MK&1rHmWEydE2rFuJLq*ls%tdkPAk2')
DB_NAME = environ.get('DB_NAME', 'eveai')
DB_PORT = environ.get('DB_PORT', '5945')
SQLALCHEMY_DATABASE_URI = f'postgresql+pg8000://{DB_USER}:{DB_PASS}@{DB_HOST}:{DB_PORT}/{DB_NAME}'
SQLALCHEMY_BINDS = {'public': SQLALCHEMY_DATABASE_URI}
EXPLAIN_TEMPLATE_LOADING = False
# Define the nginx prefix used for the specific apps
EVEAI_APP_LOCATION_PREFIX = '/admin'
EVEAI_CHAT_LOCATION_PREFIX = '/chat'
# flask-mailman settings
MAIL_USERNAME = 'eveai_super@flow-it.net'
MAIL_PASSWORD = '$6xsWGbNtx$CFMQZqc*'
# file upload settings
UPLOAD_FOLDER = '/app/tenant_files'
REDIS_USER = 'admin'
REDIS_PASS = 'b32vtDtLriSY1fL2zGrZg8IZKI0g9ucsLtVNanRFAras6oZ51wjVNB1Y05uG7uEw'
REDIS_URL = '8bciqc.stackhero-network.com'
REDIS_PORT = '9961'
REDIS_BASE_URI = f'redis://{REDIS_USER}:{REDIS_PASS}@{REDIS_URL}:{REDIS_PORT}'
# Celery settings
# eveai_app Redis Settings
CELERY_BROKER_URL = f'{REDIS_BASE_URI}/0'
CELERY_RESULT_BACKEND = f'{REDIS_BASE_URI}/0'
# eveai_chat Redis Settings
CELERY_BROKER_URL_CHAT = f'{REDIS_BASE_URI}/3'
CELERY_RESULT_BACKEND_CHAT = f'{REDIS_BASE_URI}/3'
# Session settings
SESSION_REDIS = redis.from_url(f'{REDIS_BASE_URI}/2')
# OpenAI API Keys
OPENAI_API_KEY = 'sk-proj-8R0jWzwjL7PeoPyMhJTZT3BlbkFJLb6HfRB2Hr9cEVFWEhU7'
# Groq API Keys
GROQ_API_KEY = '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'
# Anthropic API Keys
ANTHROPIC_API_KEY = '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'
# Portkey API Keys
PORTKEY_API_KEY = 'T2Dt4QTpgCvWxa1OftYCJtj7NcDZ'
# Unstructured settings
UNSTRUCTURED_API_KEY = 'pDgCrXumYhM3CNvjvwV8msMldXC3uw'
UNSTRUCTURED_BASE_URL = 'https://flowitbv-16c4us0m.api.unstructuredapp.io'
UNSTRUCTURED_FULL_URL = 'https://flowitbv-16c4us0m.api.unstructuredapp.io/general/v0/general'
# SocketIO settings
SOCKETIO_MESSAGE_QUEUE = f'{REDIS_BASE_URI}/1'
SOCKETIO_CORS_ALLOWED_ORIGINS = '*'
SOCKETIO_LOGGER = True
SOCKETIO_ENGINEIO_LOGGER = True
SOCKETIO_PING_TIMEOUT = 20000
SOCKETIO_PING_INTERVAL = 25000
SOCKETIO_MAX_IDLE_TIME = timedelta(minutes=60) # Changing this value ==> change maxConnectionDuration value in
# eveai-chat-widget.js
# Google Cloud settings
GC_PROJECT_NAME = 'eveai-420711'
GC_LOCATION = 'europe-west1'
GC_KEY_RING = 'eveai-chat'
GC_CRYPTO_KEY = 'envelope-encryption-key'
# JWT settings
JWT_SECRET_KEY = 'bsdMkmQ8ObfMD52yAFg4trrvjgjMhuIqg2fjDpD/JqvgY0ccCcmlsEnVFmR79WPiLKEA3i8a5zmejwLZKl4v9Q=='
# PATH settings
ffmpeg_path = '/usr/bin/ffmpeg'
config = {

79
config/prompts/openai/gpt-4o-mini.yaml Normal file
View File

@@ -0,0 +1,79 @@
html_parse: |
You are a top administrative assistant specialized in transforming given HTML into markdown formatted files. The generated files will be used to generate embeddings in a RAG-system.
# Best practices are:
- Respect wordings and language(s) used in the HTML.
- The following items need to be considered: headings, paragraphs, listed items (numbered or not) and tables. Images can be neglected.
- Sub-headers can be used as lists. This is true when a header is followed by a series of sub-headers without content (paragraphs or listed items). Present those sub-headers as a list.
- Be careful of encoding of the text. Everything needs to be human readable.
Process the file carefully, and take a stepped approach. The resulting markdown should be the result of the processing of the complete input html file. Answer with the pure markdown, without any other text.
HTML is between triple backquotes.
```{html}```
pdf_parse: |
You are a top administrative aid specialized in transforming given PDF-files into markdown formatted files. The generated files will be used to generate embeddings in a RAG-system.
# Best practices are:
- Respect wordings and language(s) used in the PDF.
- The following items need to be considered: headings, paragraphs, listed items (numbered or not) and tables. Images can be neglected.
- When headings are numbered, show the numbering and define the header level.
- A new item is started when a <return> is found before a full line is reached. In order to know the number of characters in a line, please check the document and the context within the document (e.g. an image could limit the number of characters temporarily).
- Paragraphs are to be stripped of newlines so they become easily readable.
- Be careful of encoding of the text. Everything needs to be human readable.
Process the file carefully, and take a stepped approach. The resulting markdown should be the result of the processing of the complete input pdf content. Answer with the pure markdown, without any other text.
PDF content is between triple backquotes.
```{pdf_content}```
summary: |
Write a concise summary of the text in {language}. The text is delimited between triple backquotes.
```{text}```
rag: |
Answer the question based on the following context, delimited between triple backquotes.
{tenant_context}
Use the following {language} in your communication, and cite the sources used.
If the question cannot be answered using the given context, say "I have insufficient information to answer this question."
Context:
```{context}```
Question:
{question}
history: |
You are a helpful assistant that details a question based on a previous context,
in such a way that the question is understandable without the previous context.
The context is a conversation history, with the HUMAN asking questions, the AI answering questions.
The history is delimited between triple backquotes.
You answer by stating the question in {language}.
History:
```{history}```
Question to be detailed:
{question}
encyclopedia: |
You have a lot of background knowledge, and as such you are some kind of
'encyclopedia' to explain general terminology. Only answer if you have a clear understanding of the question.
If not, say you do not have sufficient information to answer the question. Use the {language} in your communication.
Question:
{question}
transcript: |
You are a top administrative assistant specialized in transforming given transcriptions into markdown formatted files. The generated files will be used to generate embeddings in a RAG-system. The transcriptions originate from podcast, videos and similar material.
# Best practices and steps are:
- Respect wordings and language(s) used in the transcription. Main language is {language}.
- Sometimes, the transcript contains speech of several people participating in a conversation. Although these are not obvious from reading the file, try to detect when other people are speaking.
- Divide the transcript into several logical parts. Ensure questions and their answers are in the same logical part.
- annotate the text to identify these logical parts using headings in {language}.
- improve errors in the transcript given the context, but do not change the meaning and intentions of the transcription.
Process the file carefully, and take a stepped approach. The resulting markdown should be the result of processing the complete input transcription. Answer with the pure markdown, without any other text.
The transcript is between triple backquotes.
```{transcript}```