diff --git a/circinus/agent.py b/circinus/agent.py
index 73e9835..a837067 100644
--- a/circinus/agent.py
+++ b/circinus/agent.py
@@ -1,12 +1,60 @@
 from pathlib import Path
 
+from attrs import define
+
 from circinus.fuzzer import UserInput, candidate_prompt, fuzzing_loop
+from circinus.knowledge_base import ChunkDTO, init_knowledge_base
 from circinus.llm import GPT
 
 
+@define(hash=True)
+class Source:
+    file: str
+    page: str
+    text: str
+
+    @staticmethod
+    def curate_sources(sources: list[ChunkDTO]) -> set['Source']:
+        curated_sources = set()
+
+        for chunk in sources:
+            doc_metadata = chunk.document.doc_metadata
+
+            file_name = doc_metadata.get('file_name', '-') if doc_metadata else '-'
+            page_label = doc_metadata.get('page_label', '-') if doc_metadata else '-'
+
+            source = Source(file=file_name, page=page_label, text=chunk.text)
+            curated_sources.add(source)
+
+        return curated_sources
+
+    def __str__(self):
+        return '\n'.join([
+            f'{self.file=}',
+            f'{self.page=}',
+            f'{self.text=}'
+        ])
+
+
 class Agent:
-    def __init__(self):
+    def __init__(self, config):
         self._llm = GPT()
+        self.knowledge_base = init_knowledge_base(config)
+
+    def search_in_docs(self):
+        pass
+
+    def add_docs(self, file_name: str, file_data: str | bytes) -> None:
+        self.knowledge_base.add(file_name, file_data)
+
+    def query_docs(self, message: str):
+        return Source.curate_sources(
+            sources=self.knowledge_base.retrieve(
+                text=message,
+                limit=4,
+                prev_next_chunks=0,
+            ),
+        )
 
     def fuzz(self, documentation: str, specification: str, code: str):
         prompts = candidate_prompt(
diff --git a/circinus/knowledge_base.py b/circinus/knowledge_base.py
index db0a2e3..b938606 100644
--- a/circinus/knowledge_base.py
+++ b/circinus/knowledge_base.py
@@ -19,6 +19,12 @@
 from llama_index.storage.docstore import BaseDocumentStore
 from llama_index.storage.index_store.types import BaseIndexStore
 
+from circinus.components import (
+    doc_store_component,
+    embedding_component,
+    index_store_component,
+    llm_component,
+)
 from circinus.vector_store import ContextFilter, VectorStore
 
 
@@ -216,32 +222,11 @@ def retrieve(
         return retrieved_nodes
 
 
-def main():
-    from circinus.components import (
-        doc_store_component,
-        embedding_component,
-        index_store_component,
-        llm_component,
-    )
-    from circinus.settings import load_config
-
-    config = load_config('config.toml')
-    knowledge_base = KnowledgeBase(
+def init_knowledge_base(config) -> KnowledgeBase:
+    return KnowledgeBase(
         llm=llm_component(config=config),
         vector_store_component=VectorStore(),
         embedding=embedding_component(config=config),
         document_store=doc_store_component(),
         index_store=index_store_component(),
     )
-
-    docs = knowledge_base.add(
-        file_name='blob_documentation.txt',
-        file_data=(Path(__file__).parent / 'blob_documentation.txt').read_text(encoding='utf-8')
-    )
-
-    print(docs)
-    print(knowledge_base.retrieve('Blob.prototype.size'))
-
-
-if __name__ == '__main__':
-    main()