// good resources
// https://opensearch.org/blog/improving-document-retrieval-with-sparse-semantic-encoders/
// https://huggingface.co/opensearch-project/opensearch-neural-sparse-encoding-v1
//
// run with
// text-embeddings-router --model-id opensearch-project/opensearch-neural-sparse-encoding-v1 --pooling splade
#include
#include
#include
#include
#include
#include
#include
#include
using json = nlohmann::json;
std::vector embed(const std::vector& inputs) {
std::string url{"http://localhost:3000/embed_sparse"};
json data{
{"inputs", inputs}
};
cpr::Response r = cpr::Post(
cpr::Url{url},
cpr::Body{data.dump()},
cpr::Header{{"Content-Type", "application/json"}}
);
if (r.status_code != 200) {
throw std::runtime_error{"Bad status: " + std::to_string(r.status_code)};
}
json response = json::parse(r.text);
std::vector embeddings;
for (const auto& item : response) {
std::unordered_map map;
for (const auto& e : item) {
map.insert({e["index"], e["value"]});
}
embeddings.emplace_back(pgvector::SparseVector{map, 30522});
}
return embeddings;
}
int main() {
pqxx::connection conn{"dbname=pgvector_example"};
pqxx::nontransaction tx{conn};
tx.exec("CREATE EXTENSION IF NOT EXISTS vector");
tx.exec("DROP TABLE IF EXISTS documents");
tx.exec("CREATE TABLE documents (id bigserial PRIMARY KEY, content text, embedding sparsevec(30522))");
std::vector input{
"The dog is barking",
"The cat is purring",
"The bear is growling"
};
std::vector embeddings = embed(input);
for (size_t i = 0; i < input.size(); i++) {
tx.exec("INSERT INTO documents (content, embedding) VALUES ($1, $2)", pqxx::params{input[i], embeddings[i]});
}
std::string query{"forest"};
pgvector::SparseVector query_embedding = embed({query})[0];
pqxx::result result = tx.exec("SELECT content FROM documents ORDER BY embedding <#> $1 LIMIT 5", pqxx::params{query_embedding});
for (const auto& row : result) {
std::cout << row[0].as() << std::endl;
}
return 0;
}