Agentic RAG with iterative sub-agent search (#180)

* feat: allow agentic behavior with iterable tool calling

* feat: added system message injection to allow for smoother agentic rag

* feat: spawn subagent for search queries

* feat: first implementation of sub-agent search

* fix: make auto-retrieval assertion robust to parallel tool calls

(cherry picked from commit 2fc5d7d)

* fix: drop unsupported params in self-query LLM call for GPT-5 compatibility

(cherry picked from commit d46fcf6)

* fix: fixed token budget miscalculation

* fix: pass metadata_filter to run_tools

* fix: deduplicate chunk spans based on chunk ids

* fix: separate live messages from working messages object during sub calls

* feat: updated async_rag to match rag

* feat: code clean up

* fix: small robustness fixes for edge cases

* fix: fix required tool call in litellm

* fix: adapted tool description to make it more robust to tests

* fix: resolved pr comments

- import MetadataFilter at runtime
- remove redundant assert tool
- added clip to subagent messages
- fix bug that allowed empty tool call list to run LLM completion

---------

Co-authored-by: r-dh <remydheygere@gmail.com>

Author: MattiaMolon

src/raglite/_chatml_function_calling.py

@@ -400,7 +400,9 @@ def chatml_function_calling_with_streaming(
 
     # Case 2: Automatic or fixed tool choice
     # Case 2 step 1: Determine whether to respond with a message or a tool call
-    assert (isinstance(tool_choice, str) and tool_choice == "auto") or isinstance(tool_choice, dict)
+    assert (isinstance(tool_choice, str) and tool_choice in ("auto", "required")) or isinstance(
+        tool_choice, dict
+    )
     if isinstance(tool_choice, dict):
         tools = [t for t in tools if t["function"]["name"] == tool_choice["function"]["name"]]
         assert tools

src/raglite/_config.py

@@ -103,3 +103,4 @@ class RAGLiteConfig:
     # list[Chunk], or list[ChunkSpan].
     search_method: SearchMethod = field(default=_vector_search, compare=False)
     self_query: bool = False
+    agentic_iterations: int = 3

src/raglite/_litellm.py

@@ -36,7 +36,6 @@
 
 # Reduce the logging level for LiteLLM, flashrank, and httpx.
 litellm.suppress_debug_info = True
-litellm.drop_params = True  # Drop unsupported parameters for models like GPT-5
 os.environ["LITELLM_LOG"] = "WARNING"
 logging.getLogger("LiteLLM").setLevel(logging.WARNING)
 logging.getLogger("flashrank").setLevel(logging.WARNING)