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

Improve algorithms for HTML and PDF processing

Browse Source
This commit is contained in:
Josako
2024-07-08 15:20:45 +02:00
parent 318d23d8c6
commit ea0127b4b8
6 changed files with 176 additions and 194 deletions
Download Patch File Download Diff File Expand all files Collapse all files

78
config/config.py
View File

@@ -59,7 +59,7 @@ 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', 'openai.gpt-4-turbo', 'openai.gpt-3.5-turbo', 'mistral.mistral-large-2402']
SUPPORTED_LLMS = ['openai.gpt-4o', 'openai.gpt-4-turbo']
# Celery settings
CELERY_TASK_SERIALIZER = 'json'
@@ -69,16 +69,45 @@ class Config(object):
CELERY_ENABLE_UTC = True
# Chunk Definition, Embedding dependent
OAI_TE3S_MIN_CHUNK_SIZE = 2000
OAI_TE3S_MAX_CHUNK_SIZE = 3000
OAI_TE3L_MIN_CHUNK_SIZE = 3000
OAI_TE3L_MAX_CHUNK_SIZE = 4000
# OAI_TE3S_MIN_CHUNK_SIZE = 2000
# OAI_TE3S_MAX_CHUNK_SIZE = 3000
# OAI_TE3L_MIN_CHUNK_SIZE = 3000
# OAI_TE3L_MAX_CHUNK_SIZE = 4000
# LLM TEMPLATES
GPT4_HTML_PARSE_TEMPLATE = """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}```"""
GPT4_PDF_PARSE_TEMPLATE = """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}```
"""
GPT4_SUMMARY_TEMPLATE = """Write a concise summary of the text in {language}. The text is delimited between triple backquotes.
```{text}```"""
GPT3_5_SUMMARY_TEMPLATE = """Write a concise summary of the text in {language}. The text is delimited between triple backquotes.
```{text}```"""
GPT4_RAG_TEMPLATE = """Answer the question based on the following context, delimited between triple backquotes.
{tenant_context}
@@ -88,14 +117,6 @@ class Config(object):
```{context}```
Question:
{question}"""
GPT3_5_RAG_TEMPLATE = """Answer the question based on the following context, delimited between triple backquotes.
{tenant_context}
Use the following {language} in your communication.
If the question cannot be answered using the given context, say "I have insufficient information to answer this question."
Context:
```{context}```
Question:
{question}"""
GPT4_HISTORY_TEMPLATE = """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.
@@ -108,29 +129,12 @@ class Config(object):
Question to be detailed:
{question}"""
GPT3_5_HISTORY_TEMPLATE = """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.
{tenant_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}"""
GPT4_ENCYCLOPEDIA_TEMPLATE = """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}"""
GPT3_5_ENCYCLOPEDIA_TEMPLATE = """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}"""
GPT4_TRANSCRIPT_TEMPLATE = """You are a transcription editor that improves a given transcript on several parts
and returns markdown. Without changing what people say. The transcript is delimited between triple backquotes.
Do the following:
@@ -141,16 +145,6 @@ class Config(object):
```{transcript}```
"""
GPT3_5_TRANSCRIPT_TEMPLATE = """You are a transcription editor that improves a given transcript on several parts
and returns markdown. Without changing what people say. The transcript is delimited between triple backquotes.
Do the following:
- 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 (max 2 levels) in the same language as the transcript.
- improve errors in the transcript given the context, but leave the text intact.
```{transcript}```
"""
# SocketIO settings
# SOCKETIO_ASYNC_MODE = 'threading'
SOCKETIO_ASYNC_MODE = 'gevent'