Skip to content

Generate atomic queries instead of compound ones #24

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
robertabbott merged 1 commit into from
Feb 1, 2026
Merged

Generate atomic queries instead of compound ones #24

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
42 changes: 33 additions & 9 deletions agents/python/src/lib/chat.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,16 +52,34 @@ def GenerateDataQuestions(questions: List[DataQuestion]): # pylint: disable=inv
- 2-4 basic questions (search_effort='fast') for straightforward data lookups AND superlative/ranking queries
- 0-1 prediction market questions (search_effort='deep') about forecasts, probabilities, or future outcomes

IMPORTANT - Include superlative/ranking queries (use fast search):
CRITICAL - Generate ATOMIC, non-compound queries:
- Each query should ask for ONE metric/dimension only
- NEVER combine multiple metrics in a single query (e.g., "compare X and Y", "X vs Y", "X and Y trends")
- Instead, split compound questions into separate atomic queries

BAD (compound):
- "Compare San Francisco fentanyl consumption and population"
- "China GDP vs inflation"
- "Unemployment and wage growth in the US"

GOOD (atomic):
- "San Francisco fentanyl consumption"
- "San Francisco population"
- "China GDP"
- "China inflation"
- "US unemployment rate"
- "US wage growth"

Superlative/ranking queries are fine (these are atomic):
- "Which countries have the highest GDP per capita?"
- "Which cities have the highest rent?"
- "What are the top 10 companies by market cap?"
- "Top 10 companies by market cap"

Example:
[
{"question": "China GDP since 1960", "search_effort": "fast", "query_type": "basic"},
{"question": "China inflation rate", "search_effort": "fast", "query_type": "basic"},
{"question": "Which countries have the highest inflation rates in 2024?", "search_effort": "fast", "query_type": "basic"},
{"question": "Compare exports for east asian countries", "search_effort": "fast", "query_type": "basic"},
{"question": "What are prediction market odds for China invading Taiwan in 2025?", "search_effort": "deep", "query_type": "prediction_market"}
]
"""
Expand Down Expand Up @@ -144,18 +162,24 @@ async def chat_node(
# Build dynamic prompt based on feature toggles
if ENABLE_DEEP_QUERIES:
data_questions_instructions = """2. THEN: Use GenerateDataQuestions to create 3-6 data-focused questions with varied complexity:
- Generate ATOMIC queries - each query asks for ONE metric/dimension only
- 2-3 BASIC questions (fast search) for straightforward data: "Country X GDP 2020-2024"
- 1-2 COMPLEX questions (deep search) for analytical insights
- 0-1 PREDICTION MARKET question (deep search) if relevant: "What are prediction market odds for X in 2025?"
- Use the entities, metrics, cohorts, and time periods listed in the knowledge base context above when available
- Prefer exact entity/metric names from the knowledge base context for better search results"""
else:
data_questions_instructions = """2. THEN: Use GenerateDataQuestions to create 3-5 data-focused questions:
- 2-4 BASIC questions (fast search) for data lookups, comparisons, AND superlative/ranking queries:
* Data lookups: "Country X GDP 2020-2024"
* Superlatives: "Which cities have the highest rent?", "Which countries have the lowest unemployment?"
* Rankings: "Top 10 companies by market cap"
* Comparisons: "Compare GDP growth of X vs Y"
data_questions_instructions = """2. THEN: Use GenerateDataQuestions to create 3-6 data-focused questions:
- Generate ATOMIC queries - each query asks for ONE metric/dimension only
- Instead, split compound questions into separate atomic queries
- Examples of GOOD atomic queries:
* "US GDP 2020-2024" (single metric)
* "US inflation rate" (single metric, separate query)
* "Which countries have the highest rent?" (superlative - this is atomic)
* "Top 10 companies by revenue" (ranking - this is atomic)
- Examples of BAD compound queries to AVOID:
* "Compare US GDP and inflation" -> split into two queries
* "San Francisco population vs rent" -> split into two queries
- 0-1 PREDICTION MARKET question (deep search) if relevant: "What are prediction market odds for X in 2025?"
- Use the entities, metrics, cohorts, and time periods listed in the knowledge base context above when available
- Prefer exact entity/metric names from the knowledge base context for better search results"""
Expand Down