Anyone have a solution in their RAG app

bbig_ol_tender

Anyone have a solution in their RAG app for smart date parsing? E. G. To filter docs in a date range based on the query. I’m aware of things like spacy which can identity date entities, but then there is a missing intermediate step to go from something like “this year” to an actual date/time for a query filter.

8 comments

LLogan M

You could use an LLM to do some preprocessing on the query. Using like a pydantic program or something.

Plain Text

from typing import List
from pydanitc import BaseModel, Field

def Entity(BaseModel):
  """Represents an entity name (date, time, place, etc.) and it's associated value."""
  name: str
  value: str

def Entities(BaseModel):
  """A list of entities."""
  entities: List[Entity]


from llama_index.program import OpenAIPydanticProgram

prompt = "Given a query, extract any useful entities. If none are found, return an empty list.\nQuery: {query_str}"
program = program = OpenAIPydanticProgram.from_defaults(
    output_cls=Entities,
    prompt_template_str=prompt,
    llm=OpenAI(model="gpt-3.5-turbo-1106"),
)

entities = program(query_str="My query")
print(entities.entities)

LLogan M

You could of course tweak that as much as you need. Kind of annoying to have to call the LLM for this, but I can't think of a better way to catch those edge cases you mentioned

bbig_ol_tender

I’m thinking about fine tuning Bert for this

bbig_ol_tender

I have so so many things that this would help with

bbig_ol_tender

The prompt would be like Todays Date is YYYY-MM-DD. The user has entered the query below: xxxx
Extract the relevant date range:

LLogan M

Would have to be a generative model, not just bert. Probably T5 is a good choice 🤔

bbig_ol_tender

Yeah. I feel like this should be a pretty popular thing if done well? Rather than going full bore text to sql or w/e, no matter what your RAG stack date filters are important for narrowing the search space

LLogan M

Our auto retriever would technically handle that. Introduces an LLM step to write the filters, top k, and query

Add a reply

Find answers from the community

Anyone have a solution in their RAG app