feat: Qdrant hybrid search #2787
Conversation
Signed-off-by: Anush008 <anushshetty90@gmail.com>
Signed-off-by: Anush008 <anushshetty90@gmail.com>
|
Hey @Anush008, thanks a lot for this contribution. It was recently added in the community wishlist and I was about to get to it 😅. Really appreciate it. I'll test it from my side soon and get back!! |
Signed-off-by: Anush008 <anushshetty90@gmail.com>
a6bbc1d to
0bf5bb8
Compare
There was a problem hiding this comment.
A few comments @Anush008
For some reference on the design we follow, it'll be better for a quick reference to here- https://github.com/agno-agi/agno/blob/main/libs/agno/agno/vectordb/lancedb/lance_db.py
- We ideally would want to keep
keyword_search,vector_searchandhybrid_searchas their own functions based on the search type. - It'll be great if you could default to this approach and implement them also? Only if you get the time, else we're fine with
hybrid_searchfor now :)
Signed-off-by: Anush008 <anushshetty90@gmail.com>
…_hybrid_search() Signed-off-by: Anush008 <anushshetty90@gmail.com>
Signed-off-by: Anush008 <anushshetty90@gmail.com>
Signed-off-by: Anush008 <anushshetty90@gmail.com>
|
Hey @kausmeows. I incorporated your suggestions. |
Signed-off-by: Anush008 <anushshetty90@gmail.com>
|
Hey guys. Just bumping this PR. Please take a look when possible. |
|
Any updates? |
|
Hey @Anush008 sorry about the delay, we were working on a major feature here- #3005 on supporting knowledge based filtering (manual + agentic). There have been some major changes in the qdrant db file which i think will conflict here. Next step is hybrid search. If you'd like to resolve the conflicts and take this to the finish line please let me know? Else I can take over this PR and build on top of it!!? Thanks a lot for this amazing contribution |
…earch Signed-off-by: Anush008 <anushshetty90@gmail.com>
|
Hello @kausmeows. |
Signed-off-by: Anush008 <anushshetty90@gmail.com>
e31a2c4 to
adefab2
Compare
|
@Anush008 hybrid search is not working?? NOTE: better to add this in the cookbooks in the path- from agno.agent import Agent
from agno.knowledge.pdf_url import PDFUrlKnowledgeBase
from agno.vectordb.qdrant import Qdrant, SearchType
COLLECTION_NAME = "thai-recipes"
vector_db = Qdrant(collection=COLLECTION_NAME, url="http://localhost:6333", search_type=SearchType.hybrid)
knowledge_base = PDFUrlKnowledgeBase(
urls=["https://agno-public.s3.amazonaws.com/recipes/ThaiRecipes.pdf"],
vector_db=vector_db,
)
knowledge_base.load(recreate=False) # Comment out after first run
# Create and use the agent
agent = Agent(knowledge=knowledge_base, show_tool_calls=True)
agent.print_response(
"List down the ingredients to make Massaman Gai", markdown=True)Also change the from agno.vectordb.qdrant.qdrant import Qdrant
from agno.vectordb.search import SearchType
__all__ = [
"Qdrant", "SearchType"
] |
| key = f"meta_data.{key}" | ||
| filters = self._format_filters(filters) | ||
| if self.search_type == SearchType.vector: | ||
| results = self._run_vector_search_sync(query, limit, filters) |
There was a problem hiding this comment.
also lets rename the functions here-
if self.search_type == SearchType.vector:
results = self.vector_search(query, limit, filters)
elif self.search_type == SearchType.keyword:
results = self.keyword_search(query, limit, filters)
elif self.search_type == SearchType.hybrid:
results = self.hybrid_search(query, limit, filters) if self.search_type == SearchType.vector:
results = await self.async_vector_search(query, limit, filters)
elif self.search_type == SearchType.keyword:
results = await self.async_keyword_search(query, limit, filters)
elif self.search_type == SearchType.hybrid:
results = await self.async_hybrid_search(query, limit, filters)| from agno.agent import Agent | ||
| from agno.models.openai import OpenAIChat | ||
| from agno.tools.yfinance import YFinanceTools | ||
| from langtrace_python_sdk import langtrace # Must precede other imports |
There was a problem hiding this comment.
It has to be before the agno imports, right?
|
@Anush008 also just noticed something- Technically it is a breaking change because it requires new columns, so existing tables would have to be recreated. because my existing qdrant didn't store embeddings as wdyt, can we do that? |
Yeah. It defaults to an unnamed dense vector. But now, we're using 2 named vectors for dense and sparse.
This can be made to work, but the code becomes riddled with conditionals and will be hard to maintain. |
Hmh i see, but how big a refactor that be, i think a few conditionals won't hurt as long as they dont force users to do something that they surely will be very uncomfortable doing..? @dirkbrnd @manuhortet how do you feel about this? |
Summary
This PR adds support for hybrid searches when using Qdrant as the vector store.
qdrant/fastembed's BM25 is used for sparse embeddings by default(Customizable).
Qdrant's hybrid search reference: https://qdrant.tech/documentation/concepts/hybrid-queries/#hybrid-search
Type of change