mirror of
https://github.com/BerriAI/litellm.git
synced 2025-04-26 11:14:04 +00:00
Merge pull request #5018 from haadirakhangi/main
Qdrant Semantic Caching
This commit is contained in:
Ishaan Jaff
GitHub
commit
a34aeafdb5
5 changed files with 694 additions and 6 deletions
|
|
@ -1219,6 +1219,410 @@ class RedisSemanticCache(BaseCache):
|
|||
async def _index_info(self):
|
||||
return await self.index.ainfo()
|
||||
|
||||
class QdrantSemanticCache(BaseCache):
|
||||
def __init__(
|
||||
self,
|
||||
qdrant_url=None,
|
||||
qdrant_api_key = None,
|
||||
collection_name=None,
|
||||
similarity_threshold=None,
|
||||
quantization_config=None,
|
||||
embedding_model="text-embedding-ada-002",
|
||||
host_type = None
|
||||
):
|
||||
from litellm.llms.custom_httpx.http_handler import (
|
||||
_get_httpx_client,
|
||||
_get_async_httpx_client
|
||||
)
|
||||
|
||||
if collection_name is None:
|
||||
raise Exception("collection_name must be provided, passed None")
|
||||
|
||||
self.collection_name = collection_name
|
||||
print_verbose(
|
||||
f"qdrant semantic-cache initializing COLLECTION - {self.collection_name}"
|
||||
)
|
||||
|
||||
if similarity_threshold is None:
|
||||
raise Exception("similarity_threshold must be provided, passed None")
|
||||
self.similarity_threshold = similarity_threshold
|
||||
self.embedding_model = embedding_model
|
||||
|
||||
if host_type=="cloud":
|
||||
import os
|
||||
if qdrant_url is None:
|
||||
qdrant_url = os.getenv('QDRANT_URL')
|
||||
if qdrant_api_key is None:
|
||||
qdrant_api_key = os.getenv('QDRANT_API_KEY')
|
||||
if qdrant_url is not None and qdrant_api_key is not None:
|
||||
headers = {
|
||||
"api-key": qdrant_api_key,
|
||||
"Content-Type": "application/json"
|
||||
}
|
||||
else:
|
||||
raise Exception("Qdrant url and api_key must be provided for qdrant cloud hosting")
|
||||
elif host_type=="local":
|
||||
import os
|
||||
if qdrant_url is None:
|
||||
qdrant_url = os.getenv('QDRANT_URL')
|
||||
if qdrant_url is None:
|
||||
raise Exception("Qdrant url must be provided for qdrant local hosting")
|
||||
if qdrant_api_key is None:
|
||||
qdrant_api_key = os.getenv('QDRANT_API_KEY')
|
||||
if qdrant_api_key is None:
|
||||
print_verbose('Running locally without API Key.')
|
||||
headers= {
|
||||
"Content-Type": "application/json"
|
||||
}
|
||||
else:
|
||||
print_verbose("Running locally with API Key")
|
||||
headers = {
|
||||
"api-key": qdrant_api_key,
|
||||
"Content-Type": "application/json"
|
||||
}
|
||||
else:
|
||||
raise Exception("Host type can be either 'local' or 'cloud'")
|
||||
|
||||
self.qdrant_url = qdrant_url
|
||||
self.qdrant_api_key = qdrant_api_key
|
||||
print_verbose(f"qdrant semantic-cache qdrant_url: {self.qdrant_url}")
|
||||
|
||||
self.headers = headers
|
||||
|
||||
self.sync_client = _get_httpx_client()
|
||||
self.async_client = _get_async_httpx_client()
|
||||
|
||||
if quantization_config is None:
|
||||
print('Quantization config is not provided. Default binary quantization will be used.')
|
||||
|
||||
collection_exists = self.sync_client.get(
|
||||
url= f"{self.qdrant_url}/collections/{self.collection_name}/exists",
|
||||
headers=self.headers
|
||||
)
|
||||
if collection_exists.json()['result']['exists']:
|
||||
collection_details = self.sync_client.get(
|
||||
url=f"{self.qdrant_url}/collections/{self.collection_name}",
|
||||
headers=self.headers
|
||||
)
|
||||
self.collection_info = collection_details.json()
|
||||
print_verbose(f'Collection already exists.\nCollection details:{self.collection_info}')
|
||||
else:
|
||||
if quantization_config is None or quantization_config == 'binary':
|
||||
quantization_params = {
|
||||
"binary": {
|
||||
"always_ram": False,
|
||||
}
|
||||
}
|
||||
elif quantization_config == 'scalar':
|
||||
quantization_params = {
|
||||
"scalar": {
|
||||
"type": "int8",
|
||||
"quantile": 0.99,
|
||||
"always_ram": False
|
||||
}
|
||||
}
|
||||
elif quantization_config == 'product':
|
||||
quantization_params = {
|
||||
"product": {
|
||||
"compression": "x16",
|
||||
"always_ram": False
|
||||
}
|
||||
}
|
||||
else:
|
||||
raise Exception("Quantization config must be one of 'scalar', 'binary' or 'product'")
|
||||
|
||||
new_collection_status = self.sync_client.put(
|
||||
url=f"{self.qdrant_url}/collections/{self.collection_name}",
|
||||
json={
|
||||
"vectors": {
|
||||
"size": 1536,
|
||||
"distance": "Cosine"
|
||||
},
|
||||
"quantization_config": quantization_params
|
||||
},
|
||||
headers=self.headers
|
||||
)
|
||||
if new_collection_status.json()["result"]:
|
||||
collection_details = self.sync_client.get(
|
||||
url=f"{self.qdrant_url}/collections/{self.collection_name}",
|
||||
headers=self.headers
|
||||
)
|
||||
self.collection_info = collection_details.json()
|
||||
print_verbose(f'New collection created.\nCollection details:{self.collection_info}')
|
||||
else:
|
||||
raise Exception("Error while creating new collection")
|
||||
|
||||
def _get_cache_logic(self, cached_response: Any):
|
||||
if cached_response is None:
|
||||
return cached_response
|
||||
try:
|
||||
cached_response = json.loads(
|
||||
cached_response
|
||||
) # Convert string to dictionary
|
||||
except:
|
||||
cached_response = ast.literal_eval(cached_response)
|
||||
return cached_response
|
||||
|
||||
def set_cache(self, key, value, **kwargs):
|
||||
print_verbose(f"qdrant semantic-cache set_cache, kwargs: {kwargs}")
|
||||
import uuid
|
||||
|
||||
# get the prompt
|
||||
messages = kwargs["messages"]
|
||||
prompt = ""
|
||||
for message in messages:
|
||||
prompt += message["content"]
|
||||
|
||||
# create an embedding for prompt
|
||||
embedding_response = litellm.embedding(
|
||||
model=self.embedding_model,
|
||||
input=prompt,
|
||||
cache={"no-store": True, "no-cache": True},
|
||||
)
|
||||
|
||||
# get the embedding
|
||||
embedding = embedding_response["data"][0]["embedding"]
|
||||
|
||||
value = str(value)
|
||||
assert isinstance(value, str)
|
||||
|
||||
data = {
|
||||
"points": [
|
||||
{
|
||||
"id": str(uuid.uuid4()),
|
||||
"vector": embedding,
|
||||
"payload": {
|
||||
"text": prompt,
|
||||
"response": value,
|
||||
}
|
||||
},
|
||||
]
|
||||
}
|
||||
keys = self.sync_client.put(
|
||||
url=f"{self.qdrant_url}/collections/{self.collection_name}/points",
|
||||
headers=self.headers,
|
||||
json=data
|
||||
)
|
||||
return
|
||||
|
||||
def get_cache(self, key, **kwargs):
|
||||
print_verbose(f"sync qdrant semantic-cache get_cache, kwargs: {kwargs}")
|
||||
|
||||
# get the messages
|
||||
messages = kwargs["messages"]
|
||||
prompt = ""
|
||||
for message in messages:
|
||||
prompt += message["content"]
|
||||
|
||||
# convert to embedding
|
||||
embedding_response = litellm.embedding(
|
||||
model=self.embedding_model,
|
||||
input=prompt,
|
||||
cache={"no-store": True, "no-cache": True},
|
||||
)
|
||||
|
||||
# get the embedding
|
||||
embedding = embedding_response["data"][0]["embedding"]
|
||||
|
||||
data = {
|
||||
"vector": embedding,
|
||||
"params": {
|
||||
"quantization": {
|
||||
"ignore": False,
|
||||
"rescore": True,
|
||||
"oversampling": 3.0,
|
||||
}
|
||||
},
|
||||
"limit":1,
|
||||
"with_payload": True
|
||||
}
|
||||
|
||||
search_response = self.sync_client.post(
|
||||
url=f"{self.qdrant_url}/collections/{self.collection_name}/points/search",
|
||||
headers=self.headers,
|
||||
json=data
|
||||
)
|
||||
results = search_response.json()["result"]
|
||||
|
||||
if results == None:
|
||||
return None
|
||||
if isinstance(results, list):
|
||||
if len(results) == 0:
|
||||
return None
|
||||
|
||||
similarity = results[0]["score"]
|
||||
cached_prompt = results[0]["payload"]["text"]
|
||||
|
||||
# check similarity, if more than self.similarity_threshold, return results
|
||||
print_verbose(
|
||||
f"semantic cache: similarity threshold: {self.similarity_threshold}, similarity: {similarity}, prompt: {prompt}, closest_cached_prompt: {cached_prompt}"
|
||||
)
|
||||
if similarity >= self.similarity_threshold:
|
||||
# cache hit !
|
||||
cached_value = results[0]["payload"]["response"]
|
||||
print_verbose(
|
||||
f"got a cache hit, similarity: {similarity}, Current prompt: {prompt}, cached_prompt: {cached_prompt}"
|
||||
)
|
||||
return self._get_cache_logic(cached_response=cached_value)
|
||||
else:
|
||||
# cache miss !
|
||||
return None
|
||||
pass
|
||||
|
||||
async def async_set_cache(self, key, value, **kwargs):
|
||||
from litellm.proxy.proxy_server import llm_router, llm_model_list
|
||||
import uuid
|
||||
print_verbose(f"async qdrant semantic-cache set_cache, kwargs: {kwargs}")
|
||||
|
||||
# get the prompt
|
||||
messages = kwargs["messages"]
|
||||
prompt = ""
|
||||
for message in messages:
|
||||
prompt += message["content"]
|
||||
# create an embedding for prompt
|
||||
router_model_names = (
|
||||
[m["model_name"] for m in llm_model_list]
|
||||
if llm_model_list is not None
|
||||
else []
|
||||
)
|
||||
if llm_router is not None and self.embedding_model in router_model_names:
|
||||
user_api_key = kwargs.get("metadata", {}).get("user_api_key", "")
|
||||
embedding_response = await llm_router.aembedding(
|
||||
model=self.embedding_model,
|
||||
input=prompt,
|
||||
cache={"no-store": True, "no-cache": True},
|
||||
metadata={
|
||||
"user_api_key": user_api_key,
|
||||
"semantic-cache-embedding": True,
|
||||
"trace_id": kwargs.get("metadata", {}).get("trace_id", None),
|
||||
},
|
||||
)
|
||||
else:
|
||||
# convert to embedding
|
||||
embedding_response = await litellm.aembedding(
|
||||
model=self.embedding_model,
|
||||
input=prompt,
|
||||
cache={"no-store": True, "no-cache": True},
|
||||
)
|
||||
|
||||
# get the embedding
|
||||
embedding = embedding_response["data"][0]["embedding"]
|
||||
|
||||
value = str(value)
|
||||
assert isinstance(value, str)
|
||||
|
||||
data = {
|
||||
"points": [
|
||||
{
|
||||
"id": str(uuid.uuid4()),
|
||||
"vector": embedding,
|
||||
"payload": {
|
||||
"text": prompt,
|
||||
"response": value,
|
||||
}
|
||||
},
|
||||
]
|
||||
}
|
||||
|
||||
keys = await self.async_client.put(
|
||||
url=f"{self.qdrant_url}/collections/{self.collection_name}/points",
|
||||
headers=self.headers,
|
||||
json=data
|
||||
)
|
||||
return
|
||||
|
||||
async def async_get_cache(self, key, **kwargs):
|
||||
print_verbose(f"async qdrant semantic-cache get_cache, kwargs: {kwargs}")
|
||||
from litellm.proxy.proxy_server import llm_router, llm_model_list
|
||||
|
||||
# get the messages
|
||||
messages = kwargs["messages"]
|
||||
prompt = ""
|
||||
for message in messages:
|
||||
prompt += message["content"]
|
||||
|
||||
router_model_names = (
|
||||
[m["model_name"] for m in llm_model_list]
|
||||
if llm_model_list is not None
|
||||
else []
|
||||
)
|
||||
if llm_router is not None and self.embedding_model in router_model_names:
|
||||
user_api_key = kwargs.get("metadata", {}).get("user_api_key", "")
|
||||
embedding_response = await llm_router.aembedding(
|
||||
model=self.embedding_model,
|
||||
input=prompt,
|
||||
cache={"no-store": True, "no-cache": True},
|
||||
metadata={
|
||||
"user_api_key": user_api_key,
|
||||
"semantic-cache-embedding": True,
|
||||
"trace_id": kwargs.get("metadata", {}).get("trace_id", None),
|
||||
},
|
||||
)
|
||||
else:
|
||||
# convert to embedding
|
||||
embedding_response = await litellm.aembedding(
|
||||
model=self.embedding_model,
|
||||
input=prompt,
|
||||
cache={"no-store": True, "no-cache": True},
|
||||
)
|
||||
|
||||
# get the embedding
|
||||
embedding = embedding_response["data"][0]["embedding"]
|
||||
|
||||
data = {
|
||||
"vector": embedding,
|
||||
"params": {
|
||||
"quantization": {
|
||||
"ignore": False,
|
||||
"rescore": True,
|
||||
"oversampling": 3.0,
|
||||
}
|
||||
},
|
||||
"limit":1,
|
||||
"with_payload": True
|
||||
}
|
||||
|
||||
search_response = await self.async_client.post(
|
||||
url=f"{self.qdrant_url}/collections/{self.collection_name}/points/search",
|
||||
headers=self.headers,
|
||||
json=data
|
||||
)
|
||||
|
||||
results = search_response.json()["result"]
|
||||
|
||||
if results == None:
|
||||
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
|
||||
return None
|
||||
if isinstance(results, list):
|
||||
if len(results) == 0:
|
||||
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
|
||||
return None
|
||||
|
||||
similarity = results[0]["score"]
|
||||
cached_prompt = results[0]["payload"]["text"]
|
||||
|
||||
# check similarity, if more than self.similarity_threshold, return results
|
||||
print_verbose(
|
||||
f"semantic cache: similarity threshold: {self.similarity_threshold}, similarity: {similarity}, prompt: {prompt}, closest_cached_prompt: {cached_prompt}"
|
||||
)
|
||||
|
||||
# update kwargs["metadata"] with similarity, don't rewrite the original metadata
|
||||
kwargs.setdefault("metadata", {})["semantic-similarity"] = similarity
|
||||
|
||||
if similarity >= self.similarity_threshold:
|
||||
# cache hit !
|
||||
cached_value = results[0]["payload"]["response"]
|
||||
print_verbose(
|
||||
f"got a cache hit, similarity: {similarity}, Current prompt: {prompt}, cached_prompt: {cached_prompt}"
|
||||
)
|
||||
return self._get_cache_logic(cached_response=cached_value)
|
||||
else:
|
||||
# cache miss !
|
||||
return None
|
||||
pass
|
||||
|
||||
async def _collection_info(self):
|
||||
return self.collection_info
|
||||
|
||||
class S3Cache(BaseCache):
|
||||
def __init__(
|
||||
|
|
@ -1676,7 +2080,7 @@ class Cache:
|
|||
def __init__(
|
||||
self,
|
||||
type: Optional[
|
||||
Literal["local", "redis", "redis-semantic", "s3", "disk"]
|
||||
Literal["local", "redis", "redis-semantic", "s3", "disk", "qdrant-semantic"]
|
||||
] = "local",
|
||||
host: Optional[str] = None,
|
||||
port: Optional[str] = None,
|
||||
|
|
@ -1725,17 +2129,27 @@ class Cache:
|
|||
redis_semantic_cache_embedding_model="text-embedding-ada-002",
|
||||
redis_flush_size=None,
|
||||
disk_cache_dir=None,
|
||||
qdrant_url: Optional[str] = None,
|
||||
qdrant_api_key: Optional[str] = None,
|
||||
qdrant_collection_name: Optional[str] = None,
|
||||
qdrant_quantization_config: Optional[str] = None,
|
||||
qdrant_semantic_cache_embedding_model="text-embedding-ada-002",
|
||||
qdrant_host_type: Optional[Literal["local","cloud"]] = "local",
|
||||
**kwargs,
|
||||
):
|
||||
"""
|
||||
Initializes the cache based on the given type.
|
||||
|
||||
Args:
|
||||
type (str, optional): The type of cache to initialize. Can be "local", "redis", "redis-semantic", "s3" or "disk". Defaults to "local".
|
||||
type (str, optional): The type of cache to initialize. Can be "local", "redis", "redis-semantic", "qdrant-semantic", "s3" or "disk". Defaults to "local".
|
||||
host (str, optional): The host address for the Redis cache. Required if type is "redis".
|
||||
port (int, optional): The port number for the Redis cache. Required if type is "redis".
|
||||
password (str, optional): The password for the Redis cache. Required if type is "redis".
|
||||
similarity_threshold (float, optional): The similarity threshold for semantic-caching, Required if type is "redis-semantic"
|
||||
qdrant_url (str, optional): The url for your qdrant cluster. Required if type is "qdrant-semantic".
|
||||
qdrant_api_key (str, optional): The api_key for the local or cloud qdrant cluster. Required if qdrant_host_type is "cloud" and optional if qdrant_host_type is "local".
|
||||
qdrant_host_type (str, optional): Can be either "local" or "cloud". Should be "local" when you are running a local qdrant cluster or "cloud" when you are using a qdrant cloud cluster.
|
||||
qdrant_collection_name (str, optional): The name for your qdrant collection. Required if type is "qdrant-semantic".
|
||||
similarity_threshold (float, optional): The similarity threshold for semantic-caching, Required if type is "redis-semantic" or "qdrant-semantic".
|
||||
|
||||
supported_call_types (list, optional): List of call types to cache for. Defaults to cache == on for all call types.
|
||||
**kwargs: Additional keyword arguments for redis.Redis() cache
|
||||
|
|
@ -1760,6 +2174,16 @@ class Cache:
|
|||
embedding_model=redis_semantic_cache_embedding_model,
|
||||
**kwargs,
|
||||
)
|
||||
elif type == "qdrant-semantic":
|
||||
self.cache = QdrantSemanticCache(
|
||||
qdrant_url= qdrant_url,
|
||||
qdrant_api_key= qdrant_api_key,
|
||||
collection_name= qdrant_collection_name,
|
||||
similarity_threshold= similarity_threshold,
|
||||
quantization_config= qdrant_quantization_config,
|
||||
embedding_model= qdrant_semantic_cache_embedding_model,
|
||||
host_type=qdrant_host_type
|
||||
)
|
||||
elif type == "local":
|
||||
self.cache = InMemoryCache()
|
||||
elif type == "s3":
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue