----------------------
You can load multiple indices as a retriever by using the
QueryFusionRetriever or
RecursiveRetriever classes in LlamaIndex.
For
QueryFusionRetriever, you can fuse your indexes into a single retriever. This retriever will also generate extra queries related to the original question, and aggregate the results. Here is an example of how to use it:
from llama_index.retrievers import QueryFusionRetriever
retriever = QueryFusionRetriever(
[index_1.as_retriever(), index_2.as_retriever()],
similarity_top_k=2,
num_queries=4, # set this to 1 to disable query generation
use_async=True,
verbose=True,
# query_gen_prompt="...", # we could override the query generation prompt here
) ()
For
RecursiveRetriever, you can define a separate
IndexNode corresponding to the vector retriever for each chunk size, put all IndexNodes into a single
SummaryIndex, and then define a Recursive Retriever, with the root node being the summary index retriever. Here is an example of how to use it:
```python
from llama_index.tools import RetrieverTool
from llama_index.schema import IndexNode
from llama_index.retrievers import RecursiveRetriever
from llama_index import SummaryIndex
retriever_dict = {}
retriever_nodes = []
for chunk_size, vector_index in zip(chunk_sizes, vector_indices):
node
id = f"chunk{chunk_size}"
node = IndexNode(
text=(
"Retrieves relevant context from the Llama 2 paper (chunk size"
f" {chunk_size})"
),
index_id=node_id,
)
retriever_nodes.append(node)
retriever_dict[node_id] = vector_index.as_retriever()
summary_index = SummaryIndex(retriever_nodes)
