wehub-resource-sync
79 lines
2.7 KiB
Python
79 lines
2.7 KiB
Python
|
|
""" This example illustrates the use of a reranker model to provide a fast, effective 'in-memory' semantic
|
|
similarity, replacing an explicit retrieval using either text or embedding from a vector db.
|
|
|
|
This example uses a common pattern of trying to answer a specific factual question from a set of PDF documents.
|
|
|
|
These are the key steps of the example:
|
|
|
|
1. Sample agreements are pulled from repo, cached locally, and then iterated.
|
|
|
|
2. Each PDF document (~10-15 pages) is parsed into text chunks in memory.
|
|
-- this generates a list of dictionaries, with each dictionary entry consisting of the "text" and metadata
|
|
|
|
3. The reranker model takes as inference input both the query and the full set of text chunks.
|
|
|
|
4. The reranker model returns as output a sorted list of the (top) original text chunk dictionaries
|
|
|
|
5. The top text chunks are added as source to the prompt
|
|
|
|
6. Prompt_with_Sources run with the original query and the concatenated source to answer the question.
|
|
|
|
|
|
"""
|
|
|
|
import os
|
|
|
|
from llmware.parsers import Parser
|
|
from llmware.models import ModelCatalog
|
|
from llmware.prompts import Prompt
|
|
from llmware.setup import Setup
|
|
|
|
|
|
def rag_in_memory_with_reranker():
|
|
|
|
""" Executes a rag process in memory using semantic reranker model and bling-phi-3-gguf to answer the question. """
|
|
|
|
query = "What is the annual rate of the executive's base salary?"
|
|
|
|
sample_files_path = Setup().load_sample_files(over_write=False)
|
|
contracts_path = os.path.join(sample_files_path, "Agreements")
|
|
|
|
files = os.listdir(contracts_path)
|
|
|
|
# will use two models for the example - reranker + a 'question-answer' rag llm
|
|
reranker_model = ModelCatalog().load_model("jina-reranker-turbo")
|
|
prompter = Prompt().load_model("bling-phi-3-gguf", temperature=0.0, sample=False)
|
|
|
|
for i, doc in enumerate(files):
|
|
|
|
if doc.endswith(".pdf"):
|
|
|
|
print("\nPROCESSING: ", i, doc)
|
|
|
|
parser_output = Parser().parse_one(contracts_path,doc,save_history=False)
|
|
|
|
output = reranker_model.inference(query,parser_output,top_n=10, relevance_threshold=0.25)
|
|
|
|
use_top = 3
|
|
if len(output) > use_top:
|
|
output = output[0:use_top]
|
|
|
|
for i, results in enumerate(output):
|
|
print("semantic ranking results: ", i, results["rerank_score"], results["text"])
|
|
|
|
sources = prompter.add_source_query_results(output)
|
|
responses = prompter.prompt_with_source(query,prompt_name="default_with_context")
|
|
|
|
for i, resp in enumerate(responses):
|
|
print("\nllm answers: ", i, resp)
|
|
|
|
prompter.clear_source_materials()
|
|
|
|
return 0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
|
|
rag_in_memory_with_reranker()
|