feat: original query + better slack expansion

backend/onyx/agents/agent_search/orchestration/nodes/choose_tool.py

@@ -127,6 +127,7 @@ def choose_tool(
     override_kwargs: SearchToolOverrideKwargs = (
         force_use_tool.override_kwargs or SearchToolOverrideKwargs()
     )
+    override_kwargs.original_query = agent_config.inputs.prompt_builder.raw_user_query
 
     using_tool_calling_llm = agent_config.tooling.using_tool_calling_llm
     prompt_builder = state.prompt_snapshot or agent_config.inputs.prompt_builder

backend/onyx/context/search/federated/slack_search.py

@@ -48,8 +48,8 @@ def build_slack_queries(query: SearchQuery, llm: LLM) -> list[str]:
         time_cutoff = time_cutoff - timedelta(days=1)
         time_filter = f" after:{time_cutoff.strftime('%Y-%m-%d')}"
 
-    # use llm to generate slack queries
-    prompt = SLACK_QUERY_EXPANSION_PROMPT.format(query=query.query)
+    # use llm to generate slack queries (use original query to use same keywords as the user)
+    prompt = SLACK_QUERY_EXPANSION_PROMPT.format(query=query.original_query)
     try:
         msg = HumanMessage(content=prompt)
         response = llm.invoke([msg])

backend/onyx/context/search/models.py

@@ -154,6 +154,7 @@ class SearchRequest(ChunkContext):
     query: str
 
     expanded_queries: QueryExpansions | None = None
+    original_query: str | None = None
 
     search_type: SearchType = SearchType.SEMANTIC
 
@@ -205,6 +206,7 @@ class SearchQuery(ChunkContext):
     precomputed_query_embedding: Embedding | None = None
 
     expanded_queries: QueryExpansions | None = None
+    original_query: str | None
 
 
 class RetrievalDetails(ChunkContext):

backend/onyx/context/search/preprocessing/preprocessing.py

@@ -250,6 +250,7 @@ def retrieval_preprocessing(
 
     return SearchQuery(
         query=query,
+        original_query=search_request.original_query,
         processed_keywords=processed_keywords,
         search_type=SearchType.KEYWORD if is_keyword else SearchType.SEMANTIC,
         evaluation_type=llm_evaluation_type,

backend/onyx/prompts/federated_search.py

@@ -14,11 +14,13 @@
 1. Remove stop-words and obvious noise.
 2. Remove or down-weight meta-instructions (e.g., "show me", "summary of", "how do I") that are \
 unlikely to appear in the target messages.
-3. If the query has many keywords, produce several focused queries that keep related words together; \
+3. Stick with words used in the original query. If you need to add implied keywords (e.g., "when did" -> "date"), \
+create a separate query for it.
+4. If the query has many keywords, produce several focused queries that keep related words together; \
 never explode into single-word queries.
-4. Preserve phrases that belong together (e.g., "performance issues"); a word may appear in multiple queries.
-5. When unsure, produce both a broad and a narrow query.
-6. If the user asks for X or Y, create separate queries for X and Y.
+5. Preserve phrases that belong together (e.g., "performance issues"); a word may appear in multiple queries.
+6. When unsure, produce both a broad and a narrow query.
+7. If the user asks for X or Y, create separate queries for X and Y.
 
 Here is the original query:
 {{query}}

backend/onyx/tools/models.py

@@ -66,6 +66,7 @@ class SearchQueryInfo(BaseModel):
 
 # None indicates that the default value should be used
 class SearchToolOverrideKwargs(BaseModel):
+    original_query: str | None = None
     force_no_rerank: bool | None = None
     alternate_db_session: Session | None = None
     retrieved_sections_callback: Callable[[list[InferenceSection]], None] | None = None

backend/onyx/tools/tool_implementations/search/search_tool.py

@@ -285,6 +285,7 @@ def run(
         self, override_kwargs: SearchToolOverrideKwargs | None = None, **llm_kwargs: Any
     ) -> Generator[ToolResponse, None, None]:
         query = cast(str, llm_kwargs[QUERY_FIELD])
+        original_query = None
         precomputed_query_embedding = None
         precomputed_is_keyword = None
         precomputed_keywords = None
@@ -303,6 +304,7 @@ def run(
         kg_sources = None
         kg_chunk_id_zero_only = False
         if override_kwargs:
+            original_query = override_kwargs.original_query
             precomputed_is_keyword = override_kwargs.precomputed_is_keyword
             precomputed_keywords = override_kwargs.precomputed_keywords
             precomputed_query_embedding = override_kwargs.precomputed_query_embedding
@@ -398,6 +400,7 @@ def run(
                 precomputed_keywords=precomputed_keywords,
                 # add expanded queries
                 expanded_queries=expanded_queries,
+                original_query=original_query,
             ),
             user=self.user,
             llm=self.llm,