Skip to main content

Construct Filters

We may want to do query analysis to extract filters to pass into retrievers. One way we ask the LLM to represent these filters is as a Zod schema. There is then the issue of converting that Zod schema into a filter that can be passed into a retriever.

This can be done manually, but LangChain also provides some β€œTranslators” that are able to translate from a common syntax into filters specific to each retriever. Here, we will cover how to use those translators.

Setup​

Install dependencies​

yarn add langchain zod

In this example, year and author are both attributes to filter on.

import { z } from "zod";

const searchSchema = z.object({
query: z.string(),
startYear: z.number().optional(),
author: z.string().optional(),
});
const searchQuery: z.infer<typeof searchSchema> = {
query: "RAG",
startYear: 2022,
author: "LangChain",
};
import { Comparison, Comparator } from "langchain/chains/query_constructor/ir";

function constructComparisons(
query: z.infer<typeof searchSchema>
): Comparison[] {
const comparisons: Comparison[] = [];
if (query.startYear !== undefined) {
comparisons.push(
new Comparison("gt" as Comparator, "start_year", query.startYear)
);
}
if (query.author !== undefined) {
comparisons.push(
new Comparison("eq" as Comparator, "author", query.author)
);
}
return comparisons;
}
const comparisons = constructComparisons(searchQuery);
import { Operation, Operator } from "langchain/chains/query_constructor/ir";

const _filter = new Operation("and" as Operator, comparisons);
import { ChromaTranslator } from "langchain/retrievers/self_query/chroma";

new ChromaTranslator().visitOperation(_filter);
{
"$and": [
{ start_year: { "$gt": 2022 } },
{ author: { "$eq": "LangChain" } }
]
}

Help us out by providing feedback on this documentation page: