prompt/flow improvements

Author: joachim-danswer

backend/onyx/agents/agent_search/kb_search/nodes/a2_analyze.py

@@ -265,10 +265,7 @@ def analyze(
 Format: {output_format.value}, Broken down question: {broken_down_question}"
 
     extraction_detected_relationships = len(query_graph_relationships) > 0
-    if (
-        extraction_detected_relationships
-        or relationship_detection == KGRelationshipDetection.RELATIONSHIPS.value
-    ):
+    if extraction_detected_relationships:
         query_type = KGRelationshipDetection.RELATIONSHIPS.value
 
         if extraction_detected_relationships:

backend/onyx/agents/agent_search/kb_search/nodes/a3_generate_simple_sql.py

@@ -404,6 +404,11 @@ def generate_simple_sql(
                     "entity_table", ent_temp_view
                 )
 
+            if source_documents_sql and rel_temp_view:
+                source_documents_sql = source_documents_sql.replace(
+                    "relationship_table", rel_temp_view
+                )
+
             logger.debug(f"A3 source_documents_sql: {source_documents_sql}")
 
         scalar_result = None

backend/onyx/configs/kg_configs.py

@@ -22,10 +22,10 @@
     os.environ.get("KG_STRATEGY_GENERATION_TIMEOUT", "20")
 )
 
-KG_SQL_GENERATION_TIMEOUT: int = int(os.environ.get("KG_SQL_GENERATION_TIMEOUT", "30"))
+KG_SQL_GENERATION_TIMEOUT: int = int(os.environ.get("KG_SQL_GENERATION_TIMEOUT", "40"))
 
 KG_SQL_GENERATION_TIMEOUT_OVERRIDE: int = int(
-    os.environ.get("KG_SQL_GENERATION_TIMEOUT_OVERRIDE", "25")
+    os.environ.get("KG_SQL_GENERATION_TIMEOUT_OVERRIDE", "40")
 )
 
 KG_SQL_GENERATION_MAX_TOKENS: int = int(

backend/onyx/prompts/kg_prompts.py

@@ -41,8 +41,9 @@
 {{"entities": [<a list of entities of the prescribed entity types that you can reliably identify in the text, \
 formatted as '<ENTITY_TYPE_NAME>::<entity_name>' (please use that capitalization)>. Each entity \
 also should be followed by a list of comma-separated attribute filters for the entity, if referred to in the \
-question for that entity. Example: 'ACCOUNT::* -- [account_type: customer, status: active]' should the question be \
-'list all customer accounts', and ACCOUNT was an entity type with this attribute key/value allowed.] \
+question for that entity. CRITICAL: you can only use attributes that are mentioned above for the \
+entity type in question. Example: 'ACCOUNT::* -- [account_type: customer, status: active]' should the question be \
+'list all customer accounts', and ACCOUNT was an entity type with these attribute key/values allowed.] \
 "time_filter": <if needed, a SQL-like filter for a field called 'event_date'. Do not select anything here \
 unless you are sure that the question asks for that filter. Only apply a time_filter if the question explicitly \
 mentions a specific date, time period, or event that can be directly translated into a date filter. Do not assume \
@@ -267,12 +268,14 @@
 QUERY_ENTITY_EXTRACTION_PROMPT = f"""
 You are an expert in the area of knowledge extraction and using knowledge graphs. You are given a question \
 and asked to extract entities (with attributes if applicable) that you can reliably identify, which will then
-be matched with a known entity in the knowledge graph. You are also asked to extract time filters SHOULD \
-there be an explicit mention of a date or time frame in the QUESTION (note: last, first, etc.. DO NOT \
+be matched with a known entity in the knowledge graph. You are also asked to extract time constraints information \
+from the QUESTION. Some time constraints will be captured by entity attributes if \
+the entity type has a fitting attribute (example: 'created_at' could be a candidate for that), other times
+we will extracty an explicit time filter if no attribute fits. (Note regarding 'last', 'first', etc.: DO NOT \
 imply the need for a time filter just because the question asks for something that is not the current date. \
-They will relate to ordering that we will handle separately).
+They will relate to ordering that we will handle separately later).
 
-Today is ---today_date--- and the user asking is ---user_name---, which may or may not be relevant.
+In case useful, today is ---today_date--- and the user asking is ---user_name---, which may or may not be relevant.
 Here are the entity types that are available for extraction. Some of them may have \
 a description, others should be obvious. Also, notice that some may have attributes associated with them, which will \
 be important later.
@@ -337,6 +340,13 @@
 actual attribute may be implied.
    - don't just look at the entities that are mentioned in the question but also those that the question \
 may be about.
+  - be very careful that you only extract attributes that are listed above for the entity type in question! Do \
+not make up attributes even if they are implied! Particularly if there is a relationship type that would \
+would actually represent that information, you MUST not extract the information as an attribute. We \
+will extract the relationship type later.
+  - For the values of attributes, look at the possible values above! For example 'open' may refer to \
+'backlog', 'todo', 'in progress', etc. In cases like that construct a ';'-separated list of values that you think may fit \
+what is implied in the question (in the exanple: 'open; backlog; todo; in progress').
 
 Also, if you think the name or the title of an entity is given but name or title are not mentioned \
 explicitly as an attribute, then you should indeed extract the name/title as the entity name.
@@ -592,7 +602,11 @@
 1. SIMPLE: You think you can answer the question using a database that is aware of the entities, relationships \
 above, and is generally suitable if it is enough to either list or count entities, return dates, etc. Usually, \
 'SIMPLE' is chosen for questions of the form 'how many...' (always), or 'list the...' (often), 'when was...', \
-'what did (someone) work on...'etc.
+'what did (someone) work on...'etc. Often it is also used in cases like 'what did John work on since April?'. Here, \
+the user would expect to just see the list. So chose 'SIMPLE' here unless there are REALLY CLEAR \
+follow-up instructions for each item (like 'summarize...' , 'anaylyze...', 'what are tyhe main poinst of...'.) If \
+it is a 'what did...'-type question, chose 'SIMPLE'!
+
 2. DEEP: You think you really should ALSO leverage the actual text of sources to answer the question, which sits \
 in a vector database. Examples are 'what is discussed in...', 'summarize', 'what is the discussion about...',\
 'how does... relate to...', 'are there any mentions of... in..', 'what are the main points in...', \

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

@@ -1,3 +1,4 @@
+import copy
 import json
 from collections.abc import Callable
 from collections.abc import Generator
@@ -341,7 +342,7 @@ def run(
                 # Overwrite time-cutoff should supercede existing time-cutoff, even if defined
                 retrieval_options.filters.time_cutoff = time_cutoff
 
-        retrieval_options = retrieval_options or RetrievalDetails()
+        retrieval_options = copy.deepcopy(retrieval_options) or RetrievalDetails()
         retrieval_options.filters = retrieval_options.filters or BaseFilters()
         if kg_entities:
             retrieval_options.filters.kg_entities = kg_entities