prediction-market-agent-tooling 0.55.2.dev117__py3-none-any.whl → 0.56.0__py3-none-any.whl

prediction_market_agent_tooling/markets/seer/data_models.py

@@ -0,0 +1,27 @@
+from pydantic import BaseModel, ConfigDict, Field
+
+from prediction_market_agent_tooling.gtypes import HexBytes
+
+
+class SeerMarket(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+
+    id: HexBytes
+    title: str = Field(alias="marketName")
+    outcomes: list[str]
+    parent_market: HexBytes = Field(alias="parentMarket")
+    wrapped_tokens: list[HexBytes] = Field(alias="wrappedTokens")
+
+
+class SeerToken(BaseModel):
+    id: HexBytes
+    name: str
+    symbol: str
+
+
+class SeerPool(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+    id: HexBytes
+    liquidity: int
+    token0: SeerToken
+    token1: SeerToken

prediction_market_agent_tooling/markets/seer/seer_subgraph_handler.py

@@ -0,0 +1,142 @@
+from typing import Any
+
+from subgrounds import FieldPath
+from web3.constants import ADDRESS_ZERO
+
+from prediction_market_agent_tooling.markets.base_subgraph_handler import (
+    BaseSubgraphHandler,
+)
+from prediction_market_agent_tooling.markets.seer.data_models import (
+    SeerMarket,
+    SeerPool,
+)
+from prediction_market_agent_tooling.tools.hexbytes_custom import HexBytes
+
+INVALID_OUTCOME = "Invalid result"
+
+
+class SeerSubgraphHandler(BaseSubgraphHandler):
+    """
+    Class responsible for handling interactions with Seer subgraphs.
+    """
+
+    SEER_SUBGRAPH = "https://gateway-arbitrum.network.thegraph.com/api/{graph_api_key}/subgraphs/id/B4vyRqJaSHD8dRDb3BFRoAzuBK18c1QQcXq94JbxDxWH"
+
+    SWAPR_ALGEBRA_SUBGRAPH = "https://gateway-arbitrum.network.thegraph.com/api/{graph_api_key}/subgraphs/id/AAA1vYjxwFHzbt6qKwLHNcDSASyr1J1xVViDH8gTMFMR"
+
+    INVALID_ANSWER = "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
+
+    def __init__(self) -> None:
+        super().__init__()
+
+        self.seer_subgraph = self.sg.load_subgraph(
+            self.SEER_SUBGRAPH.format(
+                graph_api_key=self.keys.graph_api_key.get_secret_value()
+            )
+        )
+        self.swapr_algebra_subgraph = self.sg.load_subgraph(
+            self.SWAPR_ALGEBRA_SUBGRAPH.format(
+                graph_api_key=self.keys.graph_api_key.get_secret_value()
+            )
+        )
+
+    def _get_fields_for_markets(self, markets_field: FieldPath) -> list[FieldPath]:
+        fields = [
+            markets_field.id,
+            markets_field.factory,
+            markets_field.creator,
+            markets_field.marketName,
+            markets_field.outcomes,
+            markets_field.parentMarket,
+            markets_field.finalizeTs,
+            markets_field.wrappedTokens,
+        ]
+        return fields
+
+    @staticmethod
+    def filter_bicategorical_markets(markets: list[SeerMarket]) -> list[SeerMarket]:
+        # We do an extra check for the invalid outcome for safety.
+        return [
+            m for m in markets if len(m.outcomes) == 3 and INVALID_OUTCOME in m.outcomes
+        ]
+
+    @staticmethod
+    def filter_binary_markets(markets: list[SeerMarket]) -> list[SeerMarket]:
+        return [
+            market
+            for market in markets
+            if {"yes", "no"}.issubset({o.lower() for o in market.outcomes})
+        ]
+
+    @staticmethod
+    def build_filter_for_conditional_markets(
+        include_conditional_markets: bool = True,
+    ) -> dict[Any, Any]:
+        return (
+            {}
+            if include_conditional_markets
+            else {"parentMarket": ADDRESS_ZERO.lower()}
+        )
+
+    def get_bicategorical_markets(
+        self, include_conditional_markets: bool = True
+    ) -> list[SeerMarket]:
+        """Returns markets that contain 2 categories plus an invalid outcome."""
+        # Binary markets on Seer contain 3 outcomes: OutcomeA, outcomeB and an Invalid option.
+        query_filter = self.build_filter_for_conditional_markets(
+            include_conditional_markets
+        )
+        query_filter["outcomes_contains"] = [INVALID_OUTCOME]
+        markets_field = self.seer_subgraph.Query.markets(where=query_filter)
+        fields = self._get_fields_for_markets(markets_field)
+        markets = self.do_query(fields=fields, pydantic_model=SeerMarket)
+        two_category_markets = self.filter_bicategorical_markets(markets)
+        return two_category_markets
+
+    def get_binary_markets(
+        self, include_conditional_markets: bool = True
+    ) -> list[SeerMarket]:
+        two_category_markets = self.get_bicategorical_markets(
+            include_conditional_markets=include_conditional_markets
+        )
+        # Now we additionally filter markets based on YES/NO being the only outcomes.
+        binary_markets = self.filter_binary_markets(two_category_markets)
+        return binary_markets
+
+    def get_market_by_id(self, market_id: HexBytes) -> SeerMarket:
+        markets_field = self.seer_subgraph.Query.market(id=market_id.hex().lower())
+        fields = self._get_fields_for_markets(markets_field)
+        markets = self.do_query(fields=fields, pydantic_model=SeerMarket)
+        if len(markets) != 1:
+            raise ValueError(
+                f"Fetched wrong number of markets. Expected 1 but got {len(markets)}"
+            )
+        return markets[0]
+
+    def _get_fields_for_pools(self, pools_field: FieldPath) -> list[FieldPath]:
+        fields = [
+            pools_field.id,
+            pools_field.liquidity,
+            pools_field.token0.id,
+            pools_field.token0.name,
+            pools_field.token0.symbol,
+            pools_field.token1.id,
+            pools_field.token1.name,
+            pools_field.token1.symbol,
+        ]
+        return fields
+
+    def get_pools_for_market(self, market: SeerMarket) -> list[SeerPool]:
+        # We iterate through the wrapped tokens and put them in a where clause so that we hit the subgraph endpoint just once.
+        wheres = []
+        for wrapped_token in market.wrapped_tokens:
+            wheres.extend(
+                [
+                    {"token0": wrapped_token.hex().lower()},
+                    {"token1": wrapped_token.hex().lower()},
+                ]
+            )
+        pools_field = self.swapr_algebra_subgraph.Query.pools(where={"or": wheres})
+        fields = self._get_fields_for_pools(pools_field)
+        pools = self.do_query(fields=fields, pydantic_model=SeerPool)
+        return pools

prediction_market_agent_tooling/tools/caches/db_cache.py

@@ -0,0 +1,351 @@
+import hashlib
+import inspect
+import json
+from datetime import date, timedelta
+from functools import wraps
+from typing import (
+    Any,
+    Callable,
+    Sequence,
+    TypeVar,
+    cast,
+    get_args,
+    get_origin,
+    overload,
+)
+
+from pydantic import BaseModel
+from sqlalchemy import Column
+from sqlalchemy.dialects.postgresql import JSONB
+from sqlmodel import Field, Session, SQLModel, create_engine, desc, select
+
+from prediction_market_agent_tooling.config import APIKeys
+from prediction_market_agent_tooling.loggers import logger
+from prediction_market_agent_tooling.tools.datetime_utc import DatetimeUTC
+from prediction_market_agent_tooling.tools.utils import utcnow
+
+FunctionT = TypeVar("FunctionT", bound=Callable[..., Any])
+
+
+class FunctionCache(SQLModel, table=True):
+    __tablename__ = "function_cache"
+    id: int | None = Field(default=None, primary_key=True)
+    function_name: str = Field(index=True)
+    full_function_name: str = Field(index=True)
+    # Args are stored to see what was the function called with.
+    args: Any = Field(sa_column=Column(JSONB, nullable=False))
+    # Args hash is stored as a fast look-up option when looking for cache hits.
+    args_hash: str = Field(index=True)
+    result: Any = Field(sa_column=Column(JSONB, nullable=False))
+    created_at: DatetimeUTC = Field(default_factory=utcnow, index=True)
+
+
+@overload
+def db_cache(
+    func: None = None,
+    *,
+    max_age: timedelta | None = None,
+    cache_none: bool = True,
+    api_keys: APIKeys | None = None,
+    ignore_args: Sequence[str] | None = None,
+    ignore_arg_types: Sequence[type] | None = None,
+) -> Callable[[FunctionT], FunctionT]:
+    ...
+
+
+@overload
+def db_cache(
+    func: FunctionT,
+    *,
+    max_age: timedelta | None = None,
+    cache_none: bool = True,
+    api_keys: APIKeys | None = None,
+    ignore_args: Sequence[str] | None = None,
+    ignore_arg_types: Sequence[type] | None = None,
+) -> FunctionT:
+    ...
+
+
+def db_cache(
+    func: FunctionT | None = None,
+    *,
+    max_age: timedelta | None = None,
+    cache_none: bool = True,
+    api_keys: APIKeys | None = None,
+    ignore_args: Sequence[str] | None = None,
+    ignore_arg_types: Sequence[type] | None = None,
+) -> FunctionT | Callable[[FunctionT], FunctionT]:
+    if func is None:
+        # Ugly Pythonic way to support this decorator as `@postgres_cache` but also `@postgres_cache(max_age=timedelta(days=3))`
+        def decorator(func: FunctionT) -> FunctionT:
+            return db_cache(
+                func,
+                max_age=max_age,
+                cache_none=cache_none,
+                api_keys=api_keys,
+                ignore_args=ignore_args,
+                ignore_arg_types=ignore_arg_types,
+            )
+
+        return decorator
+
+    api_keys = api_keys if api_keys is not None else APIKeys()
+
+    sqlalchemy_db_url = api_keys.SQLALCHEMY_DB_URL
+    if sqlalchemy_db_url is None:
+        logger.warning(
+            f"SQLALCHEMY_DB_URL not provided in the environment, skipping function caching."
+        )
+
+    engine = (
+        create_engine(
+            sqlalchemy_db_url.get_secret_value(),
+            # Use custom json serializer and deserializer, because otherwise, for example `datetime` serialization would fail.
+            json_serializer=json_serializer,
+            json_deserializer=json_deserializer,
+        )
+        if sqlalchemy_db_url is not None
+        else None
+    )
+
+    # Create table if it doesn't exist
+    if engine is not None:
+        SQLModel.metadata.create_all(engine)
+
+    @wraps(func)
+    def wrapper(*args: Any, **kwargs: Any) -> Any:
+        # If caching is disabled, just call the function and return it
+        if not api_keys.ENABLE_CACHE:
+            return func(*args, **kwargs)
+
+        # Convert *args and **kwargs to a single dictionary, where we have names for arguments passed as args as well.
+        signature = inspect.signature(func)
+        bound_arguments = signature.bind(*args, **kwargs)
+        bound_arguments.apply_defaults()
+
+        # Convert any argument that is Pydantic model into classic dictionary, otherwise it won't be json-serializable.
+        args_dict: dict[str, Any] = bound_arguments.arguments
+
+        # Remove `self` or `cls` if present (in case of class' methods)
+        if "self" in args_dict:
+            del args_dict["self"]
+        if "cls" in args_dict:
+            del args_dict["cls"]
+
+        # Remove ignored arguments
+        if ignore_args:
+            for arg in ignore_args:
+                if arg in args_dict:
+                    del args_dict[arg]
+
+        # Remove arguments of ignored types
+        if ignore_arg_types:
+            args_dict = {
+                k: v
+                for k, v in args_dict.items()
+                if not isinstance(v, tuple(ignore_arg_types))
+            }
+
+        # Compute a hash of the function arguments used for lookup of cached results
+        arg_string = json.dumps(args_dict, sort_keys=True, default=str)
+        args_hash = hashlib.md5(arg_string.encode()).hexdigest()
+
+        # Get the full function name as concat of module and qualname, to not accidentally clash
+        full_function_name = func.__module__ + "." + func.__qualname__
+        # But also get the standard function name to easily search for it in database
+        function_name = func.__name__
+
+        # Determine if the function returns or contains Pydantic BaseModel(s)
+        return_type = func.__annotations__.get("return", None)
+        is_pydantic_model = False
+
+        if return_type is not None and contains_pydantic_model(return_type):
+            is_pydantic_model = True
+
+        # If postgres access was specified, try to find a hit
+        if engine is not None:
+            with Session(engine) as session:
+                # Try to get cached result
+                statement = (
+                    select(FunctionCache)
+                    .where(
+                        FunctionCache.function_name == function_name,
+                        FunctionCache.full_function_name == full_function_name,
+                        FunctionCache.args_hash == args_hash,
+                    )
+                    .order_by(desc(FunctionCache.created_at))
+                )
+                if max_age is not None:
+                    cutoff_time = utcnow() - max_age
+                    statement = statement.where(FunctionCache.created_at >= cutoff_time)
+                cached_result = session.exec(statement).first()
+        else:
+            cached_result = None
+
+        if cached_result:
+            logger.info(
+                # Keep the special [case-hit] identifier so we can easily track it in GCP.
+                f"[cache-hit] Cache hit for {full_function_name} with args {args_dict} and output {cached_result.result}"
+            )
+            if is_pydantic_model:
+                # If the output contains any Pydantic models, we need to initialise them.
+                try:
+                    return convert_cached_output_to_pydantic(
+                        return_type, cached_result.result
+                    )
+                except ValueError as e:
+                    # In case of backward-incompatible pydantic model, just treat it as cache miss, to not error out.
+                    logger.warning(
+                        f"Can not validate {cached_result=} into {return_type=} because {e=}, treating as cache miss."
+                    )
+                    cached_result = None
+            else:
+                return cached_result.result
+
+        # On cache miss, compute the result
+        computed_result = func(*args, **kwargs)
+        # Keep the special [case-miss] identifier so we can easily track it in GCP.
+        logger.info(
+            f"[cache-miss] Cache miss for {full_function_name} with args {args_dict}, computed the output {computed_result}"
+        )
+
+        # If postgres access was specified, save it.
+        if engine is not None and (cache_none or computed_result is not None):
+            cache_entry = FunctionCache(
+                function_name=function_name,
+                full_function_name=full_function_name,
+                args_hash=args_hash,
+                args=args_dict,
+                result=computed_result,
+                created_at=utcnow(),
+            )
+            with Session(engine) as session:
+                logger.info(f"Saving {cache_entry} into database.")
+                session.add(cache_entry)
+                session.commit()
+
+        return computed_result
+
+    return cast(FunctionT, wrapper)
+
+
+def contains_pydantic_model(return_type: Any) -> bool:
+    """
+    Check if the return type contains anything that's a Pydantic model (including nested structures, like `list[BaseModel]`, `dict[str, list[BaseModel]]`, etc.)
+    """
+    if return_type is None:
+        return False
+    origin = get_origin(return_type)
+    if origin is not None:
+        return any(contains_pydantic_model(arg) for arg in get_args(return_type))
+    if inspect.isclass(return_type):
+        return issubclass(return_type, BaseModel)
+    return False
+
+
+def json_serializer_default_fn(
+    y: DatetimeUTC | timedelta | date | BaseModel,
+) -> str | dict[str, Any]:
+    """
+    Used to serialize objects that don't support it by default into a specific string that can be deserialized out later.
+    If this function returns a dictionary, it will be called recursivelly.
+    If you add something here, also add it to `replace_custom_stringified_objects` below.
+    """
+    if isinstance(y, DatetimeUTC):
+        return f"DatetimeUTC::{y.isoformat()}"
+    elif isinstance(y, timedelta):
+        return f"timedelta::{y.total_seconds()}"
+    elif isinstance(y, date):
+        return f"date::{y.isoformat()}"
+    elif isinstance(y, BaseModel):
+        return y.model_dump()
+    raise TypeError(
+        f"Unsuported type for the default json serialize function, value is {y}."
+    )
+
+
+def json_serializer(x: Any) -> str:
+    return json.dumps(x, default=json_serializer_default_fn)
+
+
+def replace_custom_stringified_objects(obj: Any) -> Any:
+    """
+    Used to deserialize objects from `json_serializer_default_fn` into their proper form.
+    """
+    if isinstance(obj, str):
+        if obj.startswith("DatetimeUTC::"):
+            iso_str = obj[len("DatetimeUTC::") :]
+            return DatetimeUTC.to_datetime_utc(iso_str)
+        elif obj.startswith("timedelta::"):
+            total_seconds_str = obj[len("timedelta::") :]
+            return timedelta(seconds=float(total_seconds_str))
+        elif obj.startswith("date::"):
+            iso_str = obj[len("date::") :]
+            return date.fromisoformat(iso_str)
+        else:
+            return obj
+    elif isinstance(obj, dict):
+        return {k: replace_custom_stringified_objects(v) for k, v in obj.items()}
+    elif isinstance(obj, list):
+        return [replace_custom_stringified_objects(item) for item in obj]
+    else:
+        return obj
+
+
+def json_deserializer(s: str) -> Any:
+    data = json.loads(s)
+    return replace_custom_stringified_objects(data)
+
+
+def convert_cached_output_to_pydantic(return_type: Any, data: Any) -> Any:
+    """
+    Used to initialize Pydantic models from anything cached that was originally a Pydantic model in the output. Including models in nested structures.
+    """
+    # Get the origin and arguments of the model type
+    origin = get_origin(return_type)
+    args = get_args(return_type)
+
+    # Check if the data is a dictionary
+    if isinstance(data, dict):
+        # If the model has no origin, check if it is a subclass of BaseModel
+        if origin is None:
+            if inspect.isclass(return_type) and issubclass(return_type, BaseModel):
+                # Convert the dictionary to a Pydantic model
+                return return_type(
+                    **{
+                        k: convert_cached_output_to_pydantic(
+                            getattr(return_type, k, None), v
+                        )
+                        for k, v in data.items()
+                    }
+                )
+            else:
+                # If not a Pydantic model, return the data as is
+                return data
+        # If the origin is a dictionary, convert keys and values
+        elif origin is dict:
+            key_type, value_type = args
+            return {
+                convert_cached_output_to_pydantic(
+                    key_type, k
+                ): convert_cached_output_to_pydantic(value_type, v)
+                for k, v in data.items()
+            }
+        else:
+            # If the origin is not a dictionary, return the data as is
+            return data
+    # Check if the data is a list
+    elif isinstance(data, (list, tuple)):
+        # If the origin is a list or tuple, convert each item
+        if origin in {list, tuple}:
+            item_type = args[0]
+            converted_items = [
+                convert_cached_output_to_pydantic(item_type, item) for item in data
+            ]
+            return type(data)(converted_items)
+        else:
+            # If the origin is not a list or tuple, return the data as is
+            return data
+    else:
+        # If the data is neither a dictionary nor a list, return it as is
+        return data

prediction_market_agent_tooling/tools/google.py

@@ -1,11 +1,12 @@
 import typing as t
+from datetime import timedelta
 
 import tenacity
 from googleapiclient.discovery import build
 
 from prediction_market_agent_tooling.config import APIKeys
 from prediction_market_agent_tooling.loggers import logger
-from prediction_market_agent_tooling.tools.cache import persistent_inmemory_cache
+from prediction_market_agent_tooling.tools.caches.db_cache import db_cache
 
 
 @tenacity.retry(
@@ -13,7 +14,7 @@ from prediction_market_agent_tooling.tools.cache import persistent_inmemory_cach
     stop=tenacity.stop_after_attempt(3),
     after=lambda x: logger.debug(f"search_google failed, {x.attempt_number=}."),
 )
-@persistent_inmemory_cache
+@db_cache(max_age=timedelta(days=1))
 def search_google(
     query: str | None = None,
     num: int = 3,

prediction_market_agent_tooling/tools/is_invalid.py

@@ -2,7 +2,7 @@ import tenacity
 
 from prediction_market_agent_tooling.config import APIKeys
 from prediction_market_agent_tooling.loggers import logger
-from prediction_market_agent_tooling.tools.cache import persistent_inmemory_cache
+from prediction_market_agent_tooling.tools.caches.db_cache import db_cache
 from prediction_market_agent_tooling.tools.is_predictable import (
     parse_decision_yes_no_completion,
 )
@@ -34,9 +34,10 @@ QUESTION_IS_INVALID_PROMPT = """Main signs about an invalid question (sometimes
   - Which could give an incentive only to specific participants to commit an immoral violent action, but are in practice unlikely.
   - Valid: Will the US be engaged in a military conflict with a UN member state in 2021? (It’s unlikely for the US to declare war in order to win a bet on this market).
   - Valid: Will Derek Chauvin go to jail for the murder of George Flyod? (It’s unlikely that the jurors would collude to make a wrong verdict in order to win this market).
-- Questions with relative dates will resolve as invalid. Dates must be stated in absolute terms, not relative depending on the current time.
+- Questions with relative dates will resolve as invalid. Dates must be stated in absolute terms, not relative depending on the current time. But they can be relative to the event specified in the question itself.
 - Invalid: Who will be the president of the United States in 6 months? ("in 6 months depends on the current time").
 - Invalid: In the next 14 days, will Gnosis Chain gain another 1M users? ("in the next 14 days depends on the current time").
+- Valid: Will GNO price go up 10 days after Gnosis Pay cashback program is annouced? ("10 days after" is relative to the event in the question, so we can determine absolute value).
 - Questions about moral values and not facts will be resolved as invalid.
 - Invalid: "Is it ethical to eat meat?".
 
@@ -54,9 +55,9 @@ Finally, write your final decision, write `decision: ` followed by either "yes i
 """
 
 
-@persistent_inmemory_cache
 @tenacity.retry(stop=tenacity.stop_after_attempt(3), wait=tenacity.wait_fixed(1))
 @observe()
+@db_cache
 def is_invalid(
     question: str,
     engine: str = "gpt-4o",

prediction_market_agent_tooling/tools/is_predictable.py

@@ -2,7 +2,7 @@ import tenacity
 
 from prediction_market_agent_tooling.config import APIKeys
 from prediction_market_agent_tooling.loggers import logger
-from prediction_market_agent_tooling.tools.cache import persistent_inmemory_cache
+from prediction_market_agent_tooling.tools.caches.db_cache import db_cache
 from prediction_market_agent_tooling.tools.langfuse_ import (
     get_langfuse_langchain_config,
     observe,
@@ -76,9 +76,9 @@ Finally, write your final decision, write `decision: ` followed by either "yes i
 """
 
 
-@persistent_inmemory_cache
 @tenacity.retry(stop=tenacity.stop_after_attempt(3), wait=tenacity.wait_fixed(1))
 @observe()
+@db_cache
 def is_predictable_binary(
     question: str,
     engine: str = "gpt-4-1106-preview",
@@ -112,9 +112,9 @@ def is_predictable_binary(
     return parse_decision_yes_no_completion(question, completion)
 
 
-@persistent_inmemory_cache
 @tenacity.retry(stop=tenacity.stop_after_attempt(3), wait=tenacity.wait_fixed(1))
 @observe()
+@db_cache
 def is_predictable_without_description(
     question: str,
     description: str,

prediction_market_agent_tooling/tools/relevant_news_analysis/relevant_news_analysis.py

@@ -1,4 +1,4 @@
-from datetime import datetime, timedelta
+from datetime import date, timedelta
 
 from langchain_core.output_parsers import PydanticOutputParser
 from langchain_core.prompts import PromptTemplate
@@ -20,8 +20,7 @@ from prediction_market_agent_tooling.tools.relevant_news_analysis.relevant_news_
 from prediction_market_agent_tooling.tools.tavily.tavily_search import (
     get_relevant_news_since,
 )
-from prediction_market_agent_tooling.tools.tavily.tavily_storage import TavilyStorage
-from prediction_market_agent_tooling.tools.utils import check_not_none, utcnow
+from prediction_market_agent_tooling.tools.utils import check_not_none
 
 SUMMARISE_RELEVANT_NEWS_PROMPT_TEMPLATE = """
 You are an expert news analyst, tracking stories that may affect your prediction to the outcome of a particular QUESTION.
@@ -55,7 +54,7 @@ For your analysis, you should:
 def analyse_news_relevance(
     raw_content: str,
     question: str,
-    date_of_interest: datetime,
+    date_of_interest: date,
     model: str,
     temperature: float,
 ) -> RelevantNewsAnalysis:
@@ -91,19 +90,18 @@ def analyse_news_relevance(
 def get_certified_relevant_news_since(
     question: str,
     days_ago: int,
-    tavily_storage: TavilyStorage | None = None,
 ) -> RelevantNews | None:
     """
     Get relevant news since a given date for a given question. Retrieves
     possibly relevant news from tavily, then checks that it is relevant via
     an LLM call.
     """
+    news_since = date.today() - timedelta(days=days_ago)
     results = get_relevant_news_since(
         question=question,
-        days_ago=days_ago,
+        news_since=news_since,
         score_threshold=0.0,  # Be conservative to avoid missing relevant information
         max_results=3,  # A tradeoff between cost and quality. 3 seems to be a good balance.
-        tavily_storage=tavily_storage,
     )
 
     # Sort results by descending 'relevance score' to maximise the chance of
@@ -118,7 +116,7 @@ def get_certified_relevant_news_since(
         relevant_news_analysis = analyse_news_relevance(
             raw_content=check_not_none(result.raw_content),
             question=question,
-            date_of_interest=utcnow() - timedelta(days=days_ago),
+            date_of_interest=news_since,
             model="gpt-4o",  # 4o-mini isn't good enough, 1o and 1o-mini are too expensive
             temperature=0.0,
         )
@@ -140,7 +138,6 @@ def get_certified_relevant_news_since_cached(
     question: str,
     days_ago: int,
     cache: RelevantNewsResponseCache,
-    tavily_storage: TavilyStorage | None = None,
 ) -> RelevantNews | None:
     cached = cache.find(question=question, days_ago=days_ago)
 
@@ -150,7 +147,6 @@ def get_certified_relevant_news_since_cached(
         relevant_news = get_certified_relevant_news_since(
             question=question,
             days_ago=days_ago,
-            tavily_storage=tavily_storage,
         )
         cache.save(
             question=question,