Qdrant search improvements

bootstraprag/templates/llamaindex/rag_with_self_correction/self_correction_core.py

@@ -29,7 +29,8 @@ class SelfCorrectingRAG:
     ]
 
     def __init__(self, input_dir: str, similarity_top_k: int = 3, chunk_size: int = 128,
-                 chunk_overlap: int = 100, show_progress: bool = False, no_of_retries: int = 5, required_exts: list[str] = ['.pdf', '.txt']):
+                 chunk_overlap: int = 100, show_progress: bool = False, no_of_retries: int = 5,
+                 required_exts: list[str] = ['.pdf', '.txt']):
 
         self.input_dir = input_dir
         self.similarity_top_k = similarity_top_k
@@ -108,7 +109,8 @@ def query_with_retry_query_engine(self, query: str) -> RESPONSE_TYPE:
     # source nodes for the query based on llm node evaluation.
     def query_with_source_query_engine(self, query: str) -> RESPONSE_TYPE:
         retry_source_query_engine = RetrySourceQueryEngine(self.base_query_engine,
-                                                           self.query_response_evaluator)
+                                                           self.query_response_evaluator,
+                                                           max_retries=self.no_of_retries)
         retry_source_response = retry_source_query_engine.query(query)
         return retry_source_response
 
@@ -121,6 +123,7 @@ def query_with_guideline_query_engine(self, query: str) -> RESPONSE_TYPE:
                                             "The response should try to summarize where possible.\n"
         )  # just for example
         retry_guideline_query_engine = RetryGuidelineQueryEngine(self.base_query_engine,
-                                                                 guideline_eval, resynthesize_query=True)
+                                                                 guideline_eval, resynthesize_query=True,
+                                                                 max_retries=self.no_of_retries)
         retry_guideline_response = retry_guideline_query_engine.query(query)
         return retry_guideline_response

bootstraprag/templates/llamaindex/rag_with_self_correction_with_observability/self_correction_core.py

@@ -7,6 +7,7 @@
 from llama_index.llms.ollama import Ollama
 from llama_index.embeddings.ollama import OllamaEmbedding
 from llama_index.vector_stores.qdrant import QdrantVectorStore
+from llama_index. core. base. base_query_engine import BaseQueryEngine
 from llama_index.core.base.response.schema import Response, StreamingResponse, AsyncStreamingResponse, PydanticResponse
 from llama_index.core.query_engine import RetryQueryEngine, RetrySourceQueryEngine, RetryGuidelineQueryEngine
 from llama_index.core.evaluation import RelevancyEvaluator, GuidelineEvaluator
@@ -68,8 +69,8 @@ def __init__(self, input_dir: str, similarity_top_k: int = 3, chunk_size: int =
                                                                              api_key=os.environ['DB_API_KEY'])
         self.vector_store = QdrantVectorStore(client=self.client, collection_name=os.environ['COLLECTION_NAME'])
         self.query_response_evaluator = RelevancyEvaluator()
-        self.base_query_engine = None
-        self._index = None
+        self.base_query_engine: BaseQueryEngine = None
+        self._index: VectorStoreIndex = None
 
         self._load_data_and_create_engine()
 

bootstraprag/templates/qdrant/hybrid_search/.env

@@ -0,0 +1,5 @@
+DB_URL='http://localhost:6333'
+DB_API_KEY=''
+COLLECTION_NAME='YOUR_COLLECTION'
+DENSE_MODEL='sentence-transformers/all-MiniLM-L6-v2'
+SPARSE_MODEL='prithivida/Splade_PP_en_v1'

bootstraprag/templates/qdrant/hybrid_search/main.py

@@ -0,0 +1,14 @@
+import os
+
+from search import HybridSearch
+from dotenv import find_dotenv, load_dotenv
+
+_ = load_dotenv(find_dotenv())
+
+simple_search = HybridSearch(collection_name=os.environ['COLLECTION_NAME'])
+
+# uncomment if you want to index the data for the first time
+# simple_search.insert()
+result = simple_search.search(input_text='San Francisco')
+
+print(result)

bootstraprag/templates/qdrant/hybrid_search/readme.md

@@ -0,0 +1,29 @@
+## What is this project all about
+
+this is a boot straped project using bootstrap-rag cli tool. This project assume you have docker for desktop installed in your machine.
+
+### Project scafolding
+```
+.
+├── __init__.py
+├── __pycache__
+├── .env
+├── main.py
+├── readme.md
+├── requirements.txt
+└── search.py
+```
+- docker-compose.yml: if your machine does not have qdrant installed don't worry run this `docker-compose-dev.yml` in setups folder
+  - `docker-compose -f docker-compose-dev.yml up -d`
+- requirements.txt: this file has all the dependencies that a project need
+- search.py: the core logic is present in this file
+- main.py: this is the driver code to test.
+
+### How to bring in your own custom logics
+- open `search.py` and modify your `insert` and `query` functions. 
+
+or
+
+- create a `new_search_file.py` and extend it from `search.py` then override the base functionality in the new one.
+
+

bootstraprag/templates/qdrant/hybrid_search/requirements.txt

@@ -0,0 +1,3 @@
+qdrant-client==1.11.1
+python-dotenv==1.0.1
+fastembed==0.3.6

bootstraprag/templates/qdrant/hybrid_search/search.py

@@ -0,0 +1,68 @@
+import os
+import json
+from tqdm import tqdm
+from qdrant_client import qdrant_client, models
+from dotenv import find_dotenv, load_dotenv
+from qdrant_client.http.models import Distance
+from qdrant_client.conversions.common_types import QueryResponse, UpdateResult
+
+
+class HybridSearch:
+    _ = load_dotenv(find_dotenv())
+
+    def __init__(self, collection_name: str, vector_dimension: int = 384, distance: Distance = models.Distance.COSINE):
+        self.client = qdrant_client.QdrantClient(url=os.environ['DB_URL'], api_key=os.environ['DB_API_KEY'])
+
+        # set the dense and sparse embedding models
+        self.client.set_model(os.environ.get('DENSE_MODEL'))
+        self.vector_dimension = vector_dimension
+        self.distance = distance
+        self.collection_name = collection_name
+
+        self.documents = []
+        self.metadata = []
+
+        # create collection call
+        self._create_collection(collection_name=collection_name)
+
+    def _create_collection(self, collection_name):
+        if not self.client.collection_exists(collection_name=collection_name):
+            self.client.create_collection(
+                collection_name=collection_name,
+                vectors_config=self.client.get_fastembed_vector_params(),
+                sparse_vectors_config=self.client.get_fastembed_sparse_vector_params()
+            )
+
+    def _read_data(self):
+        payload_path = "data/startups-mini.json"
+
+        with open(payload_path) as fd:
+            for line in fd:
+                obj = json.loads(line)
+                self.documents.append(obj.pop("description"))
+                self.metadata.append(obj)
+
+    def insert(self) -> UpdateResult:
+        self._read_data()
+        # simple boilerplate code adjust it accordingly
+        self.client.add(
+            collection_name=self.collection_name,
+            documents=self.documents,
+            metadata=self.metadata,
+            batch_size=128, # a batch os 128 embeddings will be pushed in a single request
+            ids=tqdm(range(len(self.documents)))
+        )
+
+    def search(self, input_text: str) -> QueryResponse:
+        search_result = self.client.query(
+            collection_name=self.collection_name,
+            query_text=input_text,
+            query_filter=None,  # If you don't want any filters for now
+            limit=5,  # 5 the closest results
+        )
+        # `search_result` contains found vector ids with similarity scores
+        # along with the stored payload
+
+        # Select and return metadata
+        metadata = [hit.metadata for hit in search_result]
+        return metadata  # if not return entire search_result

bootstraprag/templates/qdrant/simple_search/.env

@@ -1,3 +1,4 @@
 DB_URL='http://localhost:6333'
 DB_API_KEY=''
 COLLECTION_NAME='YOUR_COLLECTION'
+DENSE_MODEL='sentence-transformers/all-MiniLM-L6-v2'