vk98's picture
Initial deployment of ColPali Visual Retrieval backend
a54266b
import os
import time
from typing import Any, Dict, Tuple
import asyncio
import numpy as np
import torch
from dotenv import load_dotenv
from vespa.application import Vespa
from vespa.io import VespaQueryResponse
from .colpali import SimMapGenerator
import backend.stopwords
import logging
class VespaQueryClient:
MAX_QUERY_TERMS = 64
VESPA_SCHEMA_NAME = "pdf_page"
SELECT_FIELDS = "id,title,url,blur_image,page_number,snippet,text"
def __init__(self, logger: logging.Logger):
"""
Initialize the VespaQueryClient by loading environment variables and establishing a connection to the Vespa application.
"""
load_dotenv()
self.logger = logger
if os.environ.get("USE_MTLS") == "true":
self.logger.info("Connected using mTLS")
mtls_key = os.environ.get("VESPA_CLOUD_MTLS_KEY")
mtls_cert = os.environ.get("VESPA_CLOUD_MTLS_CERT")
self.vespa_app_url = os.environ.get("VESPA_APP_MTLS_URL")
if not self.vespa_app_url:
raise ValueError(
"Please set the VESPA_APP_MTLS_URL environment variable"
)
if not mtls_cert or not mtls_key:
raise ValueError(
"USE_MTLS was true, but VESPA_CLOUD_MTLS_KEY and VESPA_CLOUD_MTLS_CERT were not set"
)
# write the key and cert to a file
mtls_key_path = "/tmp/vespa-data-plane-private-key.pem"
with open(mtls_key_path, "w") as f:
f.write(mtls_key)
mtls_cert_path = "/tmp/vespa-data-plane-public-cert.pem"
with open(mtls_cert_path, "w") as f:
f.write(mtls_cert)
# Instantiate Vespa connection
self.app = Vespa(
url=self.vespa_app_url, key=mtls_key_path, cert=mtls_cert_path
)
else:
self.logger.info("Connected using token")
self.vespa_app_url = os.environ.get("VESPA_APP_TOKEN_URL")
if not self.vespa_app_url:
raise ValueError(
"Please set the VESPA_APP_TOKEN_URL environment variable"
)
self.vespa_cloud_secret_token = os.environ.get("VESPA_CLOUD_SECRET_TOKEN")
if not self.vespa_cloud_secret_token:
raise ValueError(
"Please set the VESPA_CLOUD_SECRET_TOKEN environment variable"
)
# Instantiate Vespa connection
self.app = Vespa(
url=self.vespa_app_url,
vespa_cloud_secret_token=self.vespa_cloud_secret_token,
)
self.app.wait_for_application_up()
self.logger.info(f"Connected to Vespa at {self.vespa_app_url}")
def get_fields(self, sim_map: bool = False):
if not sim_map:
return self.SELECT_FIELDS
else:
return "summaryfeatures"
def format_query_results(
self, query: str, response: VespaQueryResponse, hits: int = 5
) -> dict:
"""
Format the Vespa query results.
Args:
query (str): The query text.
response (VespaQueryResponse): The response from Vespa.
hits (int, optional): Number of hits to display. Defaults to 5.
Returns:
dict: The JSON content of the response.
"""
query_time = response.json.get("timing", {}).get("searchtime", -1)
query_time = round(query_time, 2)
count = response.json.get("root", {}).get("fields", {}).get("totalCount", 0)
result_text = f"Query text: '{query}', query time {query_time}s, count={count}, top results:\n"
self.logger.debug(result_text)
return response.json
async def query_vespa_bm25(
self,
query: str,
q_emb: torch.Tensor,
hits: int = 3,
timeout: str = "10s",
sim_map: bool = False,
**kwargs,
) -> dict:
"""
Query Vespa using the BM25 ranking profile.
This corresponds to the "BM25" radio button in the UI.
Args:
query (str): The query text.
q_emb (torch.Tensor): Query embeddings.
hits (int, optional): Number of hits to retrieve. Defaults to 3.
timeout (str, optional): Query timeout. Defaults to "10s".
Returns:
dict: The formatted query results.
"""
async with self.app.asyncio(connections=1) as session:
query_embedding = self.format_q_embs(q_emb)
start = time.perf_counter()
response: VespaQueryResponse = await session.query(
body={
"yql": (
f"select {self.get_fields(sim_map=sim_map)} from {self.VESPA_SCHEMA_NAME} where userQuery();"
),
"ranking": self.get_rank_profile("bm25", sim_map),
"query": query,
"timeout": timeout,
"hits": hits,
"input.query(qt)": query_embedding,
"presentation.timing": True,
**kwargs,
},
)
assert response.is_successful(), response.json
stop = time.perf_counter()
self.logger.debug(
f"Query time + data transfer took: {stop - start} s, Vespa reported searchtime was "
f"{response.json.get('timing', {}).get('searchtime', -1)} s"
)
return self.format_query_results(query, response)
def float_to_binary_embedding(self, float_query_embedding: dict) -> dict:
"""
Convert float query embeddings to binary embeddings.
Args:
float_query_embedding (dict): Dictionary of float embeddings.
Returns:
dict: Dictionary of binary embeddings.
"""
binary_query_embeddings = {}
for key, vector in float_query_embedding.items():
binary_vector = (
np.packbits(np.where(np.array(vector) > 0, 1, 0))
.astype(np.int8)
.tolist()
)
binary_query_embeddings[key] = binary_vector
if len(binary_query_embeddings) >= self.MAX_QUERY_TERMS:
self.logger.warning(
f"Warning: Query has more than {self.MAX_QUERY_TERMS} terms. Truncating."
)
break
return binary_query_embeddings
def create_nn_query_strings(
self, binary_query_embeddings: dict, target_hits_per_query_tensor: int = 20
) -> Tuple[str, dict]:
"""
Create nearest neighbor query strings for Vespa.
Args:
binary_query_embeddings (dict): Binary query embeddings.
target_hits_per_query_tensor (int, optional): Target hits per query tensor. Defaults to 20.
Returns:
Tuple[str, dict]: Nearest neighbor query string and query tensor dictionary.
"""
nn_query_dict = {}
for i in range(len(binary_query_embeddings)):
nn_query_dict[f"input.query(rq{i})"] = binary_query_embeddings[i]
nn = " OR ".join(
[
f"({{targetHits:{target_hits_per_query_tensor}}}nearestNeighbor(embedding,rq{i}))"
for i in range(len(binary_query_embeddings))
]
)
return nn, nn_query_dict
def format_q_embs(self, q_embs: torch.Tensor) -> dict:
"""
Convert query embeddings to a dictionary of lists.
Args:
q_embs (torch.Tensor): Query embeddings tensor.
Returns:
dict: Dictionary where each key is an index and value is the embedding list.
"""
return {idx: emb.tolist() for idx, emb in enumerate(q_embs)}
async def get_result_from_query(
self,
query: str,
q_embs: torch.Tensor,
ranking: str,
idx_to_token: dict,
) -> Dict[str, Any]:
"""
Get query results from Vespa based on the ranking method.
Args:
query (str): The query text.
q_embs (torch.Tensor): Query embeddings.
ranking (str): The ranking method to use.
idx_to_token (dict): Index to token mapping.
Returns:
Dict[str, Any]: The query results.
"""
# Remove stopwords from the query to avoid visual emphasis on irrelevant words (e.g., "the", "and", "of")
query = backend.stopwords.filter(query)
rank_method = ranking.split("_")[0]
sim_map: bool = len(ranking.split("_")) > 1 and ranking.split("_")[1] == "sim"
if rank_method == "colpali": # ColPali
result = await self.query_vespa_colpali(
query=query, ranking=rank_method, q_emb=q_embs, sim_map=sim_map
)
elif rank_method == "hybrid": # Hybrid ColPali+BM25
result = await self.query_vespa_colpali(
query=query, ranking=rank_method, q_emb=q_embs, sim_map=sim_map
)
elif rank_method == "bm25":
result = await self.query_vespa_bm25(query, q_embs, sim_map=sim_map)
else:
raise ValueError(f"Unsupported ranking: {rank_method}")
if "root" not in result or "children" not in result["root"]:
result["root"] = {"children": []}
return result
for single_result in result["root"]["children"]:
self.logger.debug(single_result["fields"].keys())
return result
def get_sim_maps_from_query(
self, query: str, q_embs: torch.Tensor, ranking: str, idx_to_token: dict
):
"""
Get similarity maps from Vespa based on the ranking method.
Args:
query (str): The query text.
q_embs (torch.Tensor): Query embeddings.
ranking (str): The ranking method to use.
idx_to_token (dict): Index to token mapping.
Returns:
Dict[str, Any]: The query results.
"""
# Get the result by calling asyncio.run
result = asyncio.run(
self.get_result_from_query(query, q_embs, ranking, idx_to_token)
)
vespa_sim_maps = []
for single_result in result["root"]["children"]:
vespa_sim_map = single_result["fields"].get("summaryfeatures", None)
if vespa_sim_map is not None:
vespa_sim_maps.append(vespa_sim_map)
else:
raise ValueError("No sim_map found in Vespa response")
return vespa_sim_maps
async def get_full_image_from_vespa(self, doc_id: str) -> str:
"""
Retrieve the full image from Vespa for a given document ID.
Args:
doc_id (str): The document ID.
Returns:
str: The full image data.
"""
async with self.app.asyncio(connections=1) as session:
start = time.perf_counter()
response: VespaQueryResponse = await session.query(
body={
"yql": f'select full_image from {self.VESPA_SCHEMA_NAME} where id contains "{doc_id}"',
"ranking": "unranked",
"presentation.timing": True,
"ranking.matching.numThreadsPerSearch": 1,
},
)
assert response.is_successful(), response.json
stop = time.perf_counter()
self.logger.debug(
f"Getting image from Vespa took: {stop - start} s, Vespa reported searchtime was "
f"{response.json.get('timing', {}).get('searchtime', -1)} s"
)
return response.json["root"]["children"][0]["fields"]["full_image"]
def get_results_children(self, result: VespaQueryResponse) -> list:
return result["root"]["children"]
def results_to_search_results(
self, result: VespaQueryResponse, idx_to_token: dict
) -> list:
# Initialize sim_map_ fields in the result
fields_to_add = [
f"sim_map_{token}_{idx}"
for idx, token in idx_to_token.items()
if not SimMapGenerator.should_filter_token(token)
]
for child in result["root"]["children"]:
for sim_map_key in fields_to_add:
child["fields"][sim_map_key] = None
return self.get_results_children(result)
async def get_suggestions(self, query: str) -> list:
async with self.app.asyncio(connections=1) as session:
start = time.perf_counter()
yql = f'select questions from {self.VESPA_SCHEMA_NAME} where questions matches (".*{query}.*")'
response: VespaQueryResponse = await session.query(
body={
"yql": yql,
"query": query,
"ranking": "unranked",
"presentation.timing": True,
"presentation.summary": "suggestions",
"ranking.matching.numThreadsPerSearch": 1,
},
)
assert response.is_successful(), response.json
stop = time.perf_counter()
self.logger.debug(
f"Getting suggestions from Vespa took: {stop - start} s, Vespa reported searchtime was "
f"{response.json.get('timing', {}).get('searchtime', -1)} s"
)
search_results = (
response.json["root"]["children"]
if "root" in response.json and "children" in response.json["root"]
else []
)
questions = [
result["fields"]["questions"]
for result in search_results
if "questions" in result["fields"]
]
unique_questions = set([item for sublist in questions for item in sublist])
# remove an artifact from our data generation
if "string" in unique_questions:
unique_questions.remove("string")
return list(unique_questions)
def get_rank_profile(self, ranking: str, sim_map: bool) -> str:
if sim_map:
return f"{ranking}_sim"
else:
return ranking
async def query_vespa_colpali(
self,
query: str,
ranking: str,
q_emb: torch.Tensor,
target_hits_per_query_tensor: int = 100,
hnsw_explore_additional_hits: int = 300,
hits: int = 3,
timeout: str = "10s",
sim_map: bool = False,
**kwargs,
) -> dict:
"""
Query Vespa using nearest neighbor search with mixed tensors for MaxSim calculations.
This corresponds to the "ColPali" radio button in the UI.
Args:
query (str): The query text.
q_emb (torch.Tensor): Query embeddings.
target_hits_per_query_tensor (int, optional): Target hits per query tensor. Defaults to 20.
hits (int, optional): Number of hits to retrieve. Defaults to 3.
timeout (str, optional): Query timeout. Defaults to "10s".
Returns:
dict: The formatted query results.
"""
async with self.app.asyncio(connections=1) as session:
float_query_embedding = self.format_q_embs(q_emb)
binary_query_embeddings = self.float_to_binary_embedding(
float_query_embedding
)
# Mixed tensors for MaxSim calculations
query_tensors = {
"input.query(qtb)": binary_query_embeddings,
"input.query(qt)": float_query_embedding,
}
nn_string, nn_query_dict = self.create_nn_query_strings(
binary_query_embeddings, target_hits_per_query_tensor
)
query_tensors.update(nn_query_dict)
response: VespaQueryResponse = await session.query(
body={
**query_tensors,
"presentation.timing": True,
"yql": (
f"select {self.get_fields(sim_map=sim_map)} from {self.VESPA_SCHEMA_NAME} where {nn_string} or userQuery()"
),
"ranking.profile": self.get_rank_profile(
ranking=ranking, sim_map=sim_map
),
"timeout": timeout,
"hits": hits,
"query": query,
"hnsw.exploreAdditionalHits": hnsw_explore_additional_hits,
"ranking.rerankCount": 100,
**kwargs,
},
)
assert response.is_successful(), response.json
return self.format_query_results(query, response)
async def keepalive(self) -> bool:
"""
Query Vespa to keep the connection alive.
Returns:
bool: True if the connection is alive.
"""
async with self.app.asyncio(connections=1) as session:
response: VespaQueryResponse = await session.query(
body={
"yql": f"select title from {self.VESPA_SCHEMA_NAME} where true limit 1;",
"ranking": "unranked",
"query": "keepalive",
"timeout": "3s",
"hits": 1,
},
)
assert response.is_successful(), response.json
return True