apunkt/SurfSense
1
0
Fork 0
mirror of https://github.com/MODSetter/SurfSense.git synced 2026-06-06 20:15:17 +02:00
Code Issues Projects Releases Packages Wiki Activity

refactor(agents): consolidate chat runtime infra under chat/runtime

Browse source 
Move the lower-level runtime/infra modules out of multi_agent_chat/shared/
(they were never used by subagents, so they failed the shared-by-all-siblings
rule) and unify them with the already-relocated checkpointer:

  agents/runtime/                      -> agents/chat/runtime/
  mac/shared/errors.py                 -> chat/runtime/errors.py
  mac/shared/llm_config.py             -> chat/runtime/llm_config.py
  mac/shared/prompt_caching.py         -> chat/runtime/prompt_caching.py
  mac/shared/mention_resolver.py       -> chat/runtime/mention_resolver.py
  mac/shared/path_resolver.py          -> chat/runtime/path_resolver.py

These sit below the agent packages: the boundary + agent factory + shared
middleware depend on them, and they import no agent code (acyclic).
This commit is contained in:
CREDO23 2026-06-05 13:19:24 +02:00
parent 7d866a2279
commit f2a61bc0ef
52 changed files with 97 additions and 87 deletions
Download patch file Download diff file Expand all files Collapse all files

16
surfsense_backend/app/agents/chat/runtime/__init__.py Normal file
View file

@ -0,0 +1,16 @@
"""Lower-level runtime infrastructure for the chat agents.
Modules here are the foundation layer used to *run* chat agents: wired by the
boundary (routes/tasks) and/or imported by the agent factory + shared
middleware, but never part of any single agent's domain logic. Because they sit
below the agent packages, both the boundary and the agents may depend on them
(forward dependency), while they never import agent code.
Contents:
- ``checkpointer`` LangGraph Postgres checkpoint saver (boundary lifespan)
- ``llm_config`` LLM provider/model configuration resolution
- ``prompt_caching`` LiteLLM prompt-caching configuration
- ``errors`` agent-runtime error contracts (raised by MW, caught at boundary)
- ``path_resolver`` filesystem path resolution helpers
- ``mention_resolver`` @-mention resolution helpers
"""

144
surfsense_backend/app/agents/chat/runtime/checkpointer.py Normal file
View file

@ -0,0 +1,144 @@
"""
PostgreSQL-based checkpointer for LangGraph agents.
This module provides a persistent checkpointer using AsyncPostgresSaver
that stores conversation state in the PostgreSQL database.
Uses a connection pool (psycopg_pool.AsyncConnectionPool) to handle
connection lifecycle, health checks, and automatic reconnection,
preventing 'the connection is closed' errors in long-running deployments.
"""
import logging
from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
from psycopg.rows import dict_row
from psycopg_pool import AsyncConnectionPool
from app.config import config
logger = logging.getLogger(__name__)
# Global checkpointer instance (initialized lazily)
_checkpointer: AsyncPostgresSaver | None = None
_connection_pool: AsyncConnectionPool | None = None
_checkpointer_initialized: bool = False
def get_postgres_connection_string() -> str:
"""
Convert the async DATABASE_URL to a sync postgres connection string for psycopg3.
The DATABASE_URL is typically in format:
postgresql+asyncpg://user:pass@host:port/dbname
We need to convert it to:
postgresql://user:pass@host:port/dbname
"""
db_url = config.DATABASE_URL
# Handle asyncpg driver prefix
if db_url.startswith("postgresql+asyncpg://"):
return db_url.replace("postgresql+asyncpg://", "postgresql://")
# Handle other async prefixes
if "+asyncpg" in db_url:
return db_url.replace("+asyncpg", "")
return db_url
async def _create_checkpointer() -> AsyncPostgresSaver:
"""
Create a new AsyncPostgresSaver backed by a connection pool.
The connection pool automatically handles:
- Connection health checks before use
- Reconnection when connections die (idle timeout, DB restart, etc.)
- Connection lifecycle management (max_lifetime, max_idle)
"""
global _connection_pool
conn_string = get_postgres_connection_string()
_connection_pool = AsyncConnectionPool(
conninfo=conn_string,
min_size=2,
max_size=10,
# Connections are recycled after 30 minutes to avoid stale connections
max_lifetime=1800,
# Idle connections are closed after 5 minutes
max_idle=300,
open=False,
# Connection kwargs required by AsyncPostgresSaver:
# - autocommit: required for .setup() to commit checkpoint tables
# - prepare_threshold: disable prepared statements for compatibility
# - row_factory: checkpointer accesses rows as dicts (row["column"])
kwargs={
"autocommit": True,
"prepare_threshold": 0,
"row_factory": dict_row,
},
)
await _connection_pool.open(wait=True)
checkpointer = AsyncPostgresSaver(conn=_connection_pool)
logger.info("[Checkpointer] Created AsyncPostgresSaver with connection pool")
return checkpointer
async def get_checkpointer() -> AsyncPostgresSaver:
"""
Get or create the global AsyncPostgresSaver instance.
This function:
1. Creates the checkpointer with a connection pool if it doesn't exist
2. Sets up the required database tables on first call
3. Returns the cached instance on subsequent calls
The underlying connection pool handles reconnection automatically,
so a stale/closed connection will not cause OperationalError.
Returns:
AsyncPostgresSaver: The configured checkpointer instance
"""
global _checkpointer, _checkpointer_initialized
if _checkpointer is None:
_checkpointer = await _create_checkpointer()
_checkpointer_initialized = False
# Setup tables on first call (idempotent)
if not _checkpointer_initialized:
await _checkpointer.setup()
_checkpointer_initialized = True
return _checkpointer
async def setup_checkpointer_tables() -> None:
"""
Explicitly setup the checkpointer tables.
This can be called during application startup to ensure
tables exist before any agent calls.
"""
await get_checkpointer()
logger.info("[Checkpointer] PostgreSQL checkpoint tables ready")
async def close_checkpointer() -> None:
"""
Close the checkpointer connection pool.
This should be called during application shutdown.
"""
global _checkpointer, _connection_pool, _checkpointer_initialized
if _connection_pool is not None:
await _connection_pool.close()
logger.info("[Checkpointer] PostgreSQL connection pool closed")
_checkpointer = None
_connection_pool = None
_checkpointer_initialized = False

95
surfsense_backend/app/agents/chat/runtime/errors.py Normal file
View file

@ -0,0 +1,95 @@
"""
Typed error taxonomy for the SurfSense agent stack.
Used by:
- :class:`RetryAfterMiddleware` — its ``retry_on`` callable consults
the error code to decide whether a retry is appropriate.
- :class:`PermissionMiddleware` — emits ``code="permission_denied"``
errors when a deny rule trips.
- All tools — return :class:`StreamingError` payloads in
``ToolMessage.additional_kwargs["error"]`` so the model and the
retry/permission layers share a contract.
"""
from __future__ import annotations
from typing import Literal
from pydantic import BaseModel, Field
ErrorCode = Literal[
"rate_limit",
"auth",
"tool_validation",
"tool_runtime",
"context_overflow",
"provider",
"permission_denied",
"doom_loop",
"busy",
"cancelled",
]
class StreamingError(BaseModel):
"""Structured error payload attached to ``ToolMessage.additional_kwargs["error"]``.
Tools and middleware emit this so retry, permission, and routing
layers can decide what to do without parsing free-form strings.
"""
code: ErrorCode
retryable: bool = False
suggestion: str | None = None
correlation_id: str | None = None
detail: str | None = Field(
default=None,
description="Free-form additional context. Not surfaced to the model.",
)
class Config:
frozen = True
class RejectedError(Exception):
"""Raised when the user rejects a permission ask without feedback.
Caught by :class:`PermissionMiddleware`; the agent stops the current
tool fan-out and surfaces a user-facing rejection.
"""
def __init__(self, *, tool: str | None = None, pattern: str | None = None) -> None:
super().__init__(f"Permission rejected for tool {tool!r}, pattern {pattern!r}")
self.tool = tool
self.pattern = pattern
class CorrectedError(Exception):
"""Raised when the user rejects a permission ask *with* feedback.
The :class:`PermissionMiddleware` translates the feedback into a
synthetic ``ToolMessage`` so the model sees the user's correction
and can retry the request differently.
"""
def __init__(self, feedback: str, *, tool: str | None = None) -> None:
super().__init__(feedback)
self.feedback = feedback
self.tool = tool
class BusyError(Exception):
"""Raised when a second prompt arrives while the same thread is mid-stream."""
def __init__(self, request_id: str | None = None) -> None:
super().__init__("Thread is busy with another request")
self.request_id = request_id
__all__ = [
"BusyError",
"CorrectedError",
"ErrorCode",
"RejectedError",
"StreamingError",
]

624
surfsense_backend/app/agents/chat/runtime/llm_config.py Normal file
View file

@ -0,0 +1,624 @@
"""
LLM configuration utilities for SurfSense agents.
This module provides functions for loading LLM configurations from:
1. Auto mode (ID 0) - Uses LiteLLM Router for load balancing
2. YAML files (global configs with negative IDs)
3. Database NewLLMConfig table (user-created configs with positive IDs)
It also provides utilities for creating ChatLiteLLM instances and
managing prompt configurations.
"""
from collections.abc import AsyncIterator
from dataclasses import dataclass
from pathlib import Path
from typing import Any
import yaml
from langchain_core.callbacks import (
AsyncCallbackManagerForLLMRun,
CallbackManagerForLLMRun,
)
from langchain_core.messages import AIMessage, BaseMessage
from langchain_core.outputs import ChatGenerationChunk, ChatResult
from langchain_litellm import ChatLiteLLM
from litellm import get_model_info
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession
from app.agents.chat.runtime.prompt_caching import (
apply_litellm_prompt_caching,
)
from app.services.llm_router_service import (
AUTO_MODE_ID,
ChatLiteLLMRouter,
LLMRouterService,
_sanitize_content,
get_auto_mode_llm,
is_auto_mode,
)
def _sanitize_messages(messages: list[BaseMessage]) -> list[BaseMessage]:
"""Sanitize content on every message so it is safe for any provider.
Handles three cross-provider incompatibilities:
- List content with provider-specific blocks (e.g. ``thinking``)
- List content with bare strings or empty text blocks
- AI messages with empty content + tool calls: some providers (Bedrock)
convert ``""`` to ``[{"type":"text","text":""}]`` server-side then
reject the blank text. The OpenAI spec says ``content`` should be
``null`` when an assistant message only carries tool calls.
"""
for msg in messages:
if isinstance(msg.content, list):
msg.content = _sanitize_content(msg.content)
if (
isinstance(msg, AIMessage)
and (not msg.content or msg.content == "")
and getattr(msg, "tool_calls", None)
):
msg.content = None # type: ignore[assignment]
return messages
class SanitizedChatLiteLLM(ChatLiteLLM):
"""ChatLiteLLM subclass that strips provider-specific content blocks
(e.g. ``thinking`` from reasoning models) and normalises bare strings
in content arrays before forwarding to the underlying provider."""
def _generate(
self,
messages: list[BaseMessage],
stop: list[str] | None = None,
run_manager: CallbackManagerForLLMRun | None = None,
**kwargs: Any,
) -> ChatResult:
return super()._generate(
_sanitize_messages(messages), stop, run_manager, **kwargs
)
async def _astream(
self,
messages: list[BaseMessage],
stop: list[str] | None = None,
run_manager: AsyncCallbackManagerForLLMRun | None = None,
**kwargs: Any,
) -> AsyncIterator[ChatGenerationChunk]:
async for chunk in super()._astream(
_sanitize_messages(messages), stop, run_manager, **kwargs
):
yield chunk
# Provider mapping for LiteLLM model string construction.
#
# Single source of truth lives in
# :mod:`app.services.provider_capabilities` so the YAML loader (which
# runs during ``app.config`` class-body init) can resolve provider
# prefixes without dragging the agent / tools tree into module load
# order. Re-exported here under the historical ``PROVIDER_MAP`` name
# so existing callers (``llm_router_service``, ``image_gen_router_service``,
# tests) keep working unchanged.
from app.services.provider_capabilities import ( # noqa: E402
_PROVIDER_PREFIX_MAP as PROVIDER_MAP,
)
def _attach_model_profile(llm: ChatLiteLLM, model_string: str) -> None:
"""Attach a ``profile`` dict to ChatLiteLLM with model context metadata."""
try:
info = get_model_info(model_string)
max_input_tokens = info.get("max_input_tokens")
if isinstance(max_input_tokens, int) and max_input_tokens > 0:
llm.profile = {
"max_input_tokens": max_input_tokens,
"max_input_tokens_upper": max_input_tokens,
"token_count_model": model_string,
"token_count_models": [model_string],
}
except Exception:
return
@dataclass
class AgentConfig:
"""
Complete configuration for the SurfSense agent.
This combines LLM settings with prompt configuration from NewLLMConfig.
Supports Auto mode (ID 0) which uses LiteLLM Router for load balancing.
"""
# LLM Model Settings
provider: str
model_name: str
api_key: str
api_base: str | None = None
custom_provider: str | None = None
litellm_params: dict | None = None
# Prompt Configuration
system_instructions: str | None = None
use_default_system_instructions: bool = True
citations_enabled: bool = True
# Metadata
config_id: int | None = None
config_name: str | None = None
# Auto mode flag
is_auto_mode: bool = False
# Token quota and policy
billing_tier: str = "free"
is_premium: bool = False
anonymous_enabled: bool = False
quota_reserve_tokens: int | None = None
# Capability flag: best-effort True for the chat selector / catalog.
# Resolved via :func:`provider_capabilities.derive_supports_image_input`
# which prefers OpenRouter's ``architecture.input_modalities`` and
# otherwise consults LiteLLM's authoritative model map. Default True
# is the conservative-allow stance — the streaming-task safety net
# (``is_known_text_only_chat_model``) is the *only* place a False
# actually blocks a request. Setting this to False here without an
# authoritative source would silently hide vision-capable models
# (the regression we're fixing).
supports_image_input: bool = True
@classmethod
def from_auto_mode(cls) -> "AgentConfig":
"""
Create an AgentConfig for Auto mode (LiteLLM Router load balancing).
Returns:
AgentConfig instance configured for Auto mode
"""
return cls(
provider="AUTO",
model_name="auto",
api_key="", # Not needed for router
api_base=None,
custom_provider=None,
litellm_params=None,
system_instructions=None,
use_default_system_instructions=True,
citations_enabled=True,
config_id=AUTO_MODE_ID,
config_name="Auto (Fastest)",
is_auto_mode=True,
billing_tier="free",
is_premium=False,
anonymous_enabled=False,
quota_reserve_tokens=None,
# Auto routes across the configured pool, which usually
# contains at least one vision-capable deployment; the router
# will surface a 404 from a non-vision deployment as a normal
# ``allowed_fails`` event and fail over rather than blocking
# the request outright.
supports_image_input=True,
)
@classmethod
def from_new_llm_config(cls, config) -> "AgentConfig":
"""
Create an AgentConfig from a NewLLMConfig database model.
Args:
config: NewLLMConfig database model instance
Returns:
AgentConfig instance
"""
# Lazy import to avoid pulling provider_capabilities (and its
# transitive litellm import) into module-init order.
from app.services.provider_capabilities import derive_supports_image_input
provider_value = (
config.provider.value
if hasattr(config.provider, "value")
else str(config.provider)
)
litellm_params = config.litellm_params or {}
base_model = (
litellm_params.get("base_model")
if isinstance(litellm_params, dict)
else None
)
return cls(
provider=provider_value,
model_name=config.model_name,
api_key=config.api_key,
api_base=config.api_base,
custom_provider=config.custom_provider,
litellm_params=config.litellm_params,
system_instructions=config.system_instructions,
use_default_system_instructions=config.use_default_system_instructions,
citations_enabled=config.citations_enabled,
config_id=config.id,
config_name=config.name,
is_auto_mode=False,
billing_tier="free",
is_premium=False,
anonymous_enabled=False,
quota_reserve_tokens=None,
# BYOK rows have no operator-curated capability flag, so we
# ask LiteLLM (default-allow on unknown). The streaming
# safety net still blocks if the model is *explicitly*
# marked text-only.
supports_image_input=derive_supports_image_input(
provider=provider_value,
model_name=config.model_name,
base_model=base_model,
custom_provider=config.custom_provider,
),
)
@classmethod
def from_yaml_config(cls, yaml_config: dict) -> "AgentConfig":
"""
Create an AgentConfig from a YAML configuration dictionary.
YAML configs now support the same prompt configuration fields as NewLLMConfig:
- system_instructions: Custom system instructions (empty string uses defaults)
- use_default_system_instructions: Whether to use default instructions
- citations_enabled: Whether citations are enabled
Args:
yaml_config: Configuration dictionary from YAML file
Returns:
AgentConfig instance
"""
# Lazy import to avoid pulling provider_capabilities (and its
# transitive litellm import) into module-init order.
from app.services.provider_capabilities import derive_supports_image_input
# Get system instructions from YAML, default to empty string
system_instructions = yaml_config.get("system_instructions", "")
provider = yaml_config.get("provider", "").upper()
model_name = yaml_config.get("model_name", "")
custom_provider = yaml_config.get("custom_provider")
litellm_params = yaml_config.get("litellm_params") or {}
base_model = (
litellm_params.get("base_model")
if isinstance(litellm_params, dict)
else None
)
# Explicit YAML override wins; otherwise derive from LiteLLM /
# OpenRouter modalities. The YAML loader already populates this
# field, but this method is also called from
# ``load_global_llm_config_by_id``'s file fallback (hot reload),
# so we re-derive here for safety. The bool() coercion preserves
# the loader's behaviour for explicit ``true`` / ``false``
# strings that PyYAML may surface.
if "supports_image_input" in yaml_config:
supports_image_input = bool(yaml_config.get("supports_image_input"))
else:
supports_image_input = derive_supports_image_input(
provider=provider,
model_name=model_name,
base_model=base_model,
custom_provider=custom_provider,
)
return cls(
provider=provider,
model_name=model_name,
api_key=yaml_config.get("api_key", ""),
api_base=yaml_config.get("api_base"),
custom_provider=custom_provider,
litellm_params=yaml_config.get("litellm_params"),
# Prompt configuration from YAML (with defaults for backwards compatibility)
system_instructions=system_instructions if system_instructions else None,
use_default_system_instructions=yaml_config.get(
"use_default_system_instructions", True
),
citations_enabled=yaml_config.get("citations_enabled", True),
config_id=yaml_config.get("id"),
config_name=yaml_config.get("name"),
is_auto_mode=False,
billing_tier=yaml_config.get("billing_tier", "free"),
is_premium=yaml_config.get("billing_tier", "free") == "premium",
anonymous_enabled=yaml_config.get("anonymous_enabled", False),
quota_reserve_tokens=yaml_config.get("quota_reserve_tokens"),
supports_image_input=supports_image_input,
)
def load_llm_config_from_yaml(llm_config_id: int = -1) -> dict | None:
"""
Load a specific LLM config from global_llm_config.yaml.
Args:
llm_config_id: The id of the config to load (default: -1)
Returns:
LLM config dict or None if not found
"""
# Get the config file path
base_dir = Path(__file__).resolve().parent.parent.parent.parent
config_file = base_dir / "app" / "config" / "global_llm_config.yaml"
# Fallback to example file if main config doesn't exist
if not config_file.exists():
config_file = base_dir / "app" / "config" / "global_llm_config.example.yaml"
if not config_file.exists():
print("Error: No global_llm_config.yaml or example file found")
return None
try:
with open(config_file, encoding="utf-8") as f:
data = yaml.safe_load(f)
configs = data.get("global_llm_configs", [])
for cfg in configs:
if isinstance(cfg, dict) and cfg.get("id") == llm_config_id:
return cfg
print(f"Error: Global LLM config id {llm_config_id} not found")
return None
except Exception as e:
print(f"Error loading config: {e}")
return None
def load_global_llm_config_by_id(llm_config_id: int) -> dict | None:
"""
Load a global LLM config by ID, checking in-memory configs first.
This handles both static YAML configs and dynamically injected configs
(e.g. OpenRouter integration models that only exist in memory).
Args:
llm_config_id: The negative ID of the global config to load
Returns:
LLM config dict or None if not found
"""
from app.config import config as app_config
for cfg in app_config.GLOBAL_LLM_CONFIGS:
if cfg.get("id") == llm_config_id:
return cfg
# Fallback to YAML file read (covers edge cases like hot-reload)
return load_llm_config_from_yaml(llm_config_id)
async def load_new_llm_config_from_db(
session: AsyncSession,
config_id: int,
) -> "AgentConfig | None":
"""
Load a NewLLMConfig from the database by ID.
Args:
session: AsyncSession for database access
config_id: The ID of the NewLLMConfig to load
Returns:
AgentConfig instance or None if not found
"""
# Import here to avoid circular imports
from app.db import NewLLMConfig
try:
result = await session.execute(
select(NewLLMConfig).filter(NewLLMConfig.id == config_id)
)
config = result.scalars().first()
if not config:
print(f"Error: NewLLMConfig with id {config_id} not found")
return None
return AgentConfig.from_new_llm_config(config)
except Exception as e:
print(f"Error loading NewLLMConfig from database: {e}")
return None
async def load_agent_llm_config_for_search_space(
session: AsyncSession,
search_space_id: int,
) -> "AgentConfig | None":
"""
Load the agent LLM configuration for a search space.
This loads the LLM config based on the search space's agent_llm_id setting:
- Positive ID: Load from NewLLMConfig database table
- Negative ID: Load from YAML global configs
- None: Falls back to first global config (id=-1)
Args:
session: AsyncSession for database access
search_space_id: The search space ID
Returns:
AgentConfig instance or None if not found
"""
# Import here to avoid circular imports
from app.db import SearchSpace
try:
# Get the search space to check its agent_llm_id preference
result = await session.execute(
select(SearchSpace).filter(SearchSpace.id == search_space_id)
)
search_space = result.scalars().first()
if not search_space:
print(f"Error: SearchSpace with id {search_space_id} not found")
return None
# Use agent_llm_id from search space, fallback to -1 (first global config)
config_id = (
search_space.agent_llm_id if search_space.agent_llm_id is not None else -1
)
# Load the config using the unified loader
return await load_agent_config(session, config_id, search_space_id)
except Exception as e:
print(f"Error loading agent LLM config for search space {search_space_id}: {e}")
return None
async def load_agent_config(
session: AsyncSession,
config_id: int,
search_space_id: int | None = None,
) -> "AgentConfig | None":
"""
Load an agent configuration, supporting Auto mode, YAML, and database configs.
This is the main entry point for loading configurations:
- ID 0: Auto mode (uses LiteLLM Router for load balancing)
- Negative IDs: Load from YAML file (global configs)
- Positive IDs: Load from NewLLMConfig database table
Args:
session: AsyncSession for database access
config_id: The config ID (0 for Auto, negative for YAML, positive for database)
search_space_id: Optional search space ID for context
Returns:
AgentConfig instance or None if not found
"""
# Auto mode (ID 0) - use LiteLLM Router
if is_auto_mode(config_id):
if not LLMRouterService.is_initialized():
print("Error: Auto mode requested but LLM Router not initialized")
return None
return AgentConfig.from_auto_mode()
if config_id < 0:
# Check in-memory configs first (includes static YAML + dynamic OpenRouter)
from app.config import config as app_config
for cfg in app_config.GLOBAL_LLM_CONFIGS:
if cfg.get("id") == config_id:
return AgentConfig.from_yaml_config(cfg)
# Fallback to YAML file read for safety
yaml_config = load_llm_config_from_yaml(config_id)
if yaml_config:
return AgentConfig.from_yaml_config(yaml_config)
return None
else:
# Load from database (NewLLMConfig)
return await load_new_llm_config_from_db(session, config_id)
def create_chat_litellm_from_config(llm_config: dict) -> ChatLiteLLM | None:
"""
Create a ChatLiteLLM instance from a global LLM config dictionary.
Args:
llm_config: LLM configuration dictionary from YAML
Returns:
ChatLiteLLM instance or None on error
"""
# Build the model string
if llm_config.get("custom_provider"):
model_string = f"{llm_config['custom_provider']}/{llm_config['model_name']}"
else:
provider = llm_config.get("provider", "").upper()
provider_prefix = PROVIDER_MAP.get(provider, provider.lower())
model_string = f"{provider_prefix}/{llm_config['model_name']}"
# Create ChatLiteLLM instance with streaming enabled
litellm_kwargs = {
"model": model_string,
"api_key": llm_config.get("api_key"),
"streaming": True, # Enable streaming for real-time token streaming
}
# Add optional parameters
if llm_config.get("api_base"):
litellm_kwargs["api_base"] = llm_config["api_base"]
# Add any additional litellm parameters
if llm_config.get("litellm_params"):
litellm_kwargs.update(llm_config["litellm_params"])
llm = SanitizedChatLiteLLM(**litellm_kwargs)
_attach_model_profile(llm, model_string)
# Configure LiteLLM-native prompt caching (cache_control_injection_points
# for Anthropic/Bedrock/Vertex/Gemini/Azure-AI/OpenRouter/Databricks/etc.).
# ``agent_config=None`` here — the YAML path doesn't have provider intent
# in a structured form, so we set only the universal injection points.
apply_litellm_prompt_caching(llm)
return llm
def create_chat_litellm_from_agent_config(
agent_config: AgentConfig,
) -> ChatLiteLLM | ChatLiteLLMRouter | None:
"""
Create a ChatLiteLLM or ChatLiteLLMRouter instance from an AgentConfig.
For Auto mode configs, returns a ChatLiteLLMRouter that uses LiteLLM Router
for automatic load balancing across available providers.
Args:
agent_config: AgentConfig instance
Returns:
ChatLiteLLM or ChatLiteLLMRouter instance, or None on error
"""
# Handle Auto mode - return ChatLiteLLMRouter
if agent_config.is_auto_mode:
if not LLMRouterService.is_initialized():
print("Error: Auto mode requested but LLM Router not initialized")
return None
try:
router_llm = get_auto_mode_llm()
if router_llm is not None:
# Universal cache_control_injection_points only — auto-mode
# fans out across providers, so OpenAI-only kwargs (e.g.
# ``prompt_cache_key``) are left off here. ``drop_params``
# would strip them at the provider boundary anyway, but
# there's no point setting them when we don't know the
# destination.
apply_litellm_prompt_caching(router_llm, agent_config=agent_config)
return router_llm
except Exception as e:
print(f"Error creating ChatLiteLLMRouter: {e}")
return None
# Build the model string
if agent_config.custom_provider:
model_string = f"{agent_config.custom_provider}/{agent_config.model_name}"
else:
provider_prefix = PROVIDER_MAP.get(
agent_config.provider, agent_config.provider.lower()
)
model_string = f"{provider_prefix}/{agent_config.model_name}"
# Create ChatLiteLLM instance with streaming enabled
litellm_kwargs = {
"model": model_string,
"api_key": agent_config.api_key,
"streaming": True, # Enable streaming for real-time token streaming
}
# Add optional parameters
if agent_config.api_base:
litellm_kwargs["api_base"] = agent_config.api_base
# Add any additional litellm parameters
if agent_config.litellm_params:
litellm_kwargs.update(agent_config.litellm_params)
llm = SanitizedChatLiteLLM(**litellm_kwargs)
_attach_model_profile(llm, model_string)
# Build-time prompt caching: sets ``cache_control_injection_points`` for
# all providers and (for OpenAI/DeepSeek/xAI) ``prompt_cache_retention``.
# Per-thread ``prompt_cache_key`` is layered on later in
# ``create_surfsense_deep_agent`` once ``thread_id`` is known.
apply_litellm_prompt_caching(llm, agent_config=agent_config)
return llm

277
surfsense_backend/app/agents/chat/runtime/mention_resolver.py Normal file
View file

@ -0,0 +1,277 @@
"""Resolve @-mention chips to canonical virtual paths and substitute the
user-visible ``@title`` tokens with backtick-wrapped paths in the prompt
the agent sees.
The frontend's mention seam is a single discriminated-union list of
``{kind: "doc" | "folder", id, title, document_type?}`` chips (see
``surfsense_web/atoms/chat/mentioned-documents.atom.ts``). When a turn
reaches the backend stream task we have three needs that this module
centralises:
1. Map each chip to its canonical virtual path
(``/documents/.../file.xml`` for docs, ``/documents/MyFolder/`` for
folders) so the agent sees concrete filesystem locations instead of
ambiguous ``@``-titles.
2. Substitute ``@title`` tokens in the user-typed text with backtick-
wrapped paths so the path becomes part of the ``HumanMessage`` body
the LLM consumes — without rewriting the persisted user message
text (which keeps ``@title`` so chip rendering on reload is
unchanged).
3. Surface the resolved id sets (docs + folders) to the priority
middleware so it can render ``[USER-MENTIONED]`` priority entries
without re-doing path resolution.
This is intentionally one module — see the architectural note in
``mention-paths-and-folders`` plan: previously the doc-resolution lived
inline in ``stream_new_chat`` and the folder mention had no resolution
at all. Centralising both behind a single ``resolve_mentions`` call
turns a leaky multi-field seam into a single deeper interface.
"""
from __future__ import annotations
import logging
from dataclasses import dataclass, field
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession
from app.agents.chat.runtime.path_resolver import (
DOCUMENTS_ROOT,
build_path_index,
doc_to_virtual_path,
)
from app.db import Document, Folder
from app.schemas.new_chat import MentionedDocumentInfo
logger = logging.getLogger(__name__)
@dataclass(frozen=True)
class ResolvedMention:
"""Canonical view of a single @-mention chip.
``virtual_path`` is the path the agent will see (no trailing slash
for documents, trailing ``/`` for folders to match the convention
used by ``KnowledgeTreeMiddleware``).
"""
kind: str # "doc" | "folder"
id: int
title: str
virtual_path: str
@dataclass
class ResolvedMentionSet:
"""Aggregate result of resolving a turn's mention chips.
``token_to_path`` maps ``@title`` (the literal token the user typed
and the editor emitted) to the canonical virtual path for that
chip. It is produced longest-token-first so substitution mirrors
``parseMentionSegments`` on the frontend (a longer title like
``@Project Roadmap`` is never shadowed by a shorter prefix
``@Project``).
``mentioned_document_ids`` is an ordered, deduped list consumed by
the priority middleware downstream — see
``KnowledgePriorityMiddleware._compute_priority_paths``.
"""
mentions: list[ResolvedMention] = field(default_factory=list)
token_to_path: list[tuple[str, str]] = field(default_factory=list)
mentioned_document_ids: list[int] = field(default_factory=list)
mentioned_folder_ids: list[int] = field(default_factory=list)
def _folder_virtual_path(folder_id: int, folder_paths: dict[int, str]) -> str:
"""Return ``/documents/Folder/Sub/`` for a folder id.
Falls back to the documents root when the folder is missing from
the index (deleted or in a different search space). Trailing slash
matches ``KnowledgeTreeMiddleware`` (``/documents/MyFolder/``) so
the agent's ``ls`` can dispatch on it as a directory.
"""
base = folder_paths.get(folder_id, DOCUMENTS_ROOT)
return f"{base}/" if not base.endswith("/") else base
async def resolve_mentions(
session: AsyncSession,
*,
search_space_id: int,
mentioned_documents: list[MentionedDocumentInfo] | None,
mentioned_document_ids: list[int] | None = None,
mentioned_folder_ids: list[int] | None = None,
) -> ResolvedMentionSet:
"""Resolve every @-mention chip on a turn into virtual paths.
The function takes both the ``mentioned_documents`` discriminated
list (chip metadata used for substitution + persistence) and the
parallel id arrays (``mentioned_document_ids``,
``mentioned_folder_ids``) for two reasons:
* Legacy clients that haven't migrated to the unified chip list
still send the id arrays — we treat the union as authoritative.
* The id arrays are the canonical input to
``KnowledgePriorityMiddleware`` (via ``SurfSenseContextSchema``);
returning the deduped, validated lists lets the route forward
them unchanged.
Resolution is best-effort: a chip whose id no longer exists (e.g.
document was deleted between mention and submit) is silently
dropped. The agent still sees the user's original text, just
without a backtick-path substitution for that chip.
"""
chip_doc_ids: list[int] = []
chip_folder_ids: list[int] = []
chip_titles_by_id: dict[tuple[str, int], str] = {}
if mentioned_documents:
for chip in mentioned_documents:
kind = chip.kind
if kind == "folder":
chip_folder_ids.append(chip.id)
elif kind == "doc":
chip_doc_ids.append(chip.id)
chip_titles_by_id[(kind, chip.id)] = chip.title
doc_id_pool: list[int] = list(
dict.fromkeys(
[
*(mentioned_document_ids or []),
*chip_doc_ids,
]
)
)
folder_id_pool: list[int] = list(
dict.fromkeys([*(mentioned_folder_ids or []), *chip_folder_ids])
)
if not doc_id_pool and not folder_id_pool:
return ResolvedMentionSet()
index = await build_path_index(session, search_space_id)
doc_rows: dict[int, Document] = {}
if doc_id_pool:
result = await session.execute(
select(Document).where(
Document.search_space_id == search_space_id,
Document.id.in_(doc_id_pool),
)
)
for row in result.scalars().all():
doc_rows[row.id] = row
folder_rows: dict[int, Folder] = {}
if folder_id_pool:
result = await session.execute(
select(Folder).where(
Folder.search_space_id == search_space_id,
Folder.id.in_(folder_id_pool),
)
)
for row in result.scalars().all():
folder_rows[row.id] = row
resolved: list[ResolvedMention] = []
accepted_doc_ids: list[int] = []
accepted_folder_ids: list[int] = []
for doc_id in doc_id_pool:
row = doc_rows.get(doc_id)
if row is None:
logger.debug(
"mention_resolver: dropping doc id=%s (not found in space=%s)",
doc_id,
search_space_id,
)
continue
title = chip_titles_by_id.get(("doc", doc_id), str(row.title or ""))
path = doc_to_virtual_path(
doc_id=row.id,
title=str(row.title or "untitled"),
folder_id=row.folder_id,
index=index,
)
resolved.append(
ResolvedMention(kind="doc", id=row.id, title=title, virtual_path=path)
)
accepted_doc_ids.append(row.id)
for folder_id in folder_id_pool:
row = folder_rows.get(folder_id)
if row is None:
logger.debug(
"mention_resolver: dropping folder id=%s (not found in space=%s)",
folder_id,
search_space_id,
)
continue
title = chip_titles_by_id.get(("folder", folder_id), str(row.name or ""))
path = _folder_virtual_path(row.id, index.folder_paths)
resolved.append(
ResolvedMention(kind="folder", id=row.id, title=title, virtual_path=path)
)
accepted_folder_ids.append(row.id)
token_to_path: list[tuple[str, str]] = []
seen_tokens: set[str] = set()
for mention in resolved:
if not mention.title:
continue
token = f"@{mention.title}"
if token in seen_tokens:
continue
seen_tokens.add(token)
token_to_path.append((token, mention.virtual_path))
token_to_path.sort(key=lambda pair: len(pair[0]), reverse=True)
return ResolvedMentionSet(
mentions=resolved,
token_to_path=token_to_path,
mentioned_document_ids=accepted_doc_ids,
mentioned_folder_ids=accepted_folder_ids,
)
def substitute_in_text(text: str, token_to_path: list[tuple[str, str]]) -> str:
"""Replace each ``@title`` token with a backtick-wrapped virtual path.
Mirrors ``parseMentionSegments`` on the frontend: longest token
first, single forward pass, no regex (titles can contain regex
metacharacters). The substitution is idempotent for already-
substituted text because the backtick-wrapped path no longer
starts with ``@``.
Empty / no-op cases short-circuit so callers can pass this through
unconditionally without paying for a scan.
"""
if not text or not token_to_path:
return text
out: list[str] = []
i = 0
n = len(text)
while i < n:
matched: tuple[str, str] | None = None
for token, path in token_to_path:
if text.startswith(token, i):
matched = (token, path)
break
if matched is None:
out.append(text[i])
i += 1
continue
token, path = matched
out.append(f"`{path}`")
i += len(token)
return "".join(out)
__all__ = [
"ResolvedMention",
"ResolvedMentionSet",
"resolve_mentions",
"substitute_in_text",
]

351
surfsense_backend/app/agents/chat/runtime/path_resolver.py Normal file
View file

@ -0,0 +1,351 @@
"""Canonical virtual-path resolver for SurfSense knowledge-base documents.
This module is the single source of truth for mapping ``Document`` rows to
virtual paths under ``/documents/`` and back. It is used by:
* :class:`KnowledgeTreeMiddleware` (rendering the workspace tree)
* :class:`KnowledgePriorityMiddleware` (computing priority paths)
* :class:`KBPostgresBackend` (``als_info`` / ``aread`` / move operations)
* :class:`KnowledgeBasePersistenceMiddleware` (resolving moves and creates)
Centralising the logic ensures that title-collision suffixes, folder paths,
and ``unique_identifier_hash`` lookups never drift between renders and
commits.
"""
from __future__ import annotations
import re
from dataclasses import dataclass, field
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession
from app.db import Document, DocumentType, Folder
from app.utils.document_converters import generate_unique_identifier_hash
DOCUMENTS_ROOT = "/documents"
"""Root virtual folder for all KB documents."""
_INVALID_FILENAME_CHARS = re.compile(r"[\\/:*?\"<>|]+")
_WHITESPACE_RUN = re.compile(r"\s+")
def safe_filename(value: str, *, fallback: str = "untitled.xml") -> str:
"""Convert arbitrary text into a filesystem-safe ``.xml`` filename."""
name = _INVALID_FILENAME_CHARS.sub("_", value).strip()
name = _WHITESPACE_RUN.sub(" ", name)
if not name:
name = fallback
if len(name) > 180:
name = name[:180].rstrip()
if not name.lower().endswith(".xml"):
name = f"{name}.xml"
return name
def safe_folder_segment(value: str, *, fallback: str = "folder") -> str:
"""Sanitize a single folder name into a path-safe segment."""
name = _INVALID_FILENAME_CHARS.sub("_", value).strip()
name = _WHITESPACE_RUN.sub(" ", name)
if not name:
return fallback
if len(name) > 180:
name = name[:180].rstrip()
return name
def _suffix_with_doc_id(filename: str, doc_id: int | None) -> str:
if doc_id is None:
return filename
if not filename.lower().endswith(".xml"):
return f"{filename} ({doc_id}).xml"
stem = filename[:-4]
return f"{stem} ({doc_id}).xml"
_SUFFIX_PATTERN = re.compile(r"\s\((\d+)\)\.xml$", re.IGNORECASE)
def parse_doc_id_suffix(filename: str) -> tuple[str, int | None]:
"""Strip a trailing ``" (<doc_id>).xml"`` suffix; return ``(stem, doc_id)``.
If no suffix is present, returns ``(stem_without_xml_extension, None)``.
"""
match = _SUFFIX_PATTERN.search(filename)
if match:
doc_id = int(match.group(1))
stem = filename[: match.start()]
return stem, doc_id
if filename.lower().endswith(".xml"):
return filename[:-4], None
return filename, None
@dataclass
class PathIndex:
"""In-memory occupancy snapshot used by :func:`doc_to_virtual_path`.
Built once per call site so collision handling is deterministic and so
we don't perform N folder lookups per render.
"""
folder_paths: dict[int, str] = field(default_factory=dict)
"""``Folder.id`` -> absolute virtual folder path under ``/documents``."""
occupants: dict[str, int] = field(default_factory=dict)
"""virtual path -> ``Document.id`` already occupying that path (this render)."""
async def _build_folder_paths(
session: AsyncSession,
search_space_id: int,
) -> dict[int, str]:
"""Compute ``Folder.id`` -> absolute virtual path under ``/documents``."""
result = await session.execute(
select(Folder.id, Folder.name, Folder.parent_id).where(
Folder.search_space_id == search_space_id
)
)
rows = result.all()
by_id = {row.id: {"name": row.name, "parent_id": row.parent_id} for row in rows}
cache: dict[int, str] = {}
def resolve(folder_id: int) -> str:
if folder_id in cache:
return cache[folder_id]
parts: list[str] = []
cursor: int | None = folder_id
visited: set[int] = set()
while cursor is not None and cursor in by_id and cursor not in visited:
visited.add(cursor)
entry = by_id[cursor]
parts.append(safe_folder_segment(str(entry["name"])))
cursor = entry["parent_id"]
parts.reverse()
path = f"{DOCUMENTS_ROOT}/" + "/".join(parts) if parts else DOCUMENTS_ROOT
cache[folder_id] = path
return path
for folder_id in by_id:
resolve(folder_id)
return cache
async def build_path_index(
session: AsyncSession,
search_space_id: int,
*,
populate_occupants: bool = True,
) -> PathIndex:
"""Build a :class:`PathIndex` for a search space.
``populate_occupants`` controls whether the occupancy map is pre-seeded
from existing ``Document`` rows. Most callers want this so that
:func:`doc_to_virtual_path` can detect collisions across the whole space;
the persistence middleware sets this to ``False`` when it is iterating to
decide where to place fresh documents.
"""
folder_paths = await _build_folder_paths(session, search_space_id)
occupants: dict[str, int] = {}
if populate_occupants:
rows = await session.execute(
select(Document.id, Document.title, Document.folder_id).where(
Document.search_space_id == search_space_id,
)
)
for row in rows.all():
base = folder_paths.get(row.folder_id, DOCUMENTS_ROOT)
filename = safe_filename(str(row.title or "untitled"))
path = f"{base}/{filename}"
if path in occupants and occupants[path] != row.id:
path = f"{base}/{_suffix_with_doc_id(filename, row.id)}"
occupants[path] = row.id
return PathIndex(folder_paths=folder_paths, occupants=occupants)
def doc_to_virtual_path(
*,
doc_id: int | None,
title: str,
folder_id: int | None,
index: PathIndex,
) -> str:
"""Return the canonical virtual path for a document.
Mutates ``index.occupants`` so subsequent calls see this assignment and
deterministically pick a different suffix for the next colliding doc.
"""
base = index.folder_paths.get(folder_id, DOCUMENTS_ROOT)
filename = safe_filename(str(title or "untitled"))
path = f"{base}/{filename}"
occupant = index.occupants.get(path)
if occupant is not None and occupant != doc_id:
path = f"{base}/{_suffix_with_doc_id(filename, doc_id)}"
if doc_id is not None:
index.occupants[path] = doc_id
return path
async def virtual_path_to_doc(
session: AsyncSession,
*,
search_space_id: int,
virtual_path: str,
) -> Document | None:
"""Resolve a virtual path back to a ``Document`` row.
Resolution order:
1. ``Document.unique_identifier_hash`` lookup (fast path for paths created
by SurfSense itself — every NOTE write goes through this hash).
2. If the basename carries a ``" (<doc_id>).xml"`` disambiguation suffix,
try a direct id lookup constrained to the search space.
3. Title-from-basename + folder-resolution lookup as a last resort.
"""
if not virtual_path or not virtual_path.startswith(DOCUMENTS_ROOT):
return None
unique_hash = generate_unique_identifier_hash(
DocumentType.NOTE,
virtual_path,
search_space_id,
)
result = await session.execute(
select(Document).where(
Document.search_space_id == search_space_id,
Document.unique_identifier_hash == unique_hash,
)
)
document = result.scalar_one_or_none()
if document is not None:
return document
rel = virtual_path[len(DOCUMENTS_ROOT) :].lstrip("/")
if not rel:
return None
parts = [p for p in rel.split("/") if p]
if not parts:
return None
basename = parts[-1]
folder_parts = parts[:-1]
stem, suffix_doc_id = parse_doc_id_suffix(basename)
if suffix_doc_id is not None:
result = await session.execute(
select(Document).where(
Document.search_space_id == search_space_id,
Document.id == suffix_doc_id,
)
)
document = result.scalar_one_or_none()
if document is not None:
return document
folder_id = await _resolve_folder_id(
session, search_space_id=search_space_id, folder_parts=folder_parts
)
title_candidates: list[str] = []
raw_title = stem
title_candidates.append(raw_title)
if raw_title.endswith(".xml"):
title_candidates.append(raw_title[:-4])
for candidate in dict.fromkeys(title_candidates):
if not candidate:
continue
query = select(Document).where(
Document.search_space_id == search_space_id,
Document.title == candidate,
)
if folder_id is None:
query = query.where(Document.folder_id.is_(None))
else:
query = query.where(Document.folder_id == folder_id)
result = await session.execute(query)
document = result.scalars().first()
if document is not None:
return document
# Fallback: title-as-string lookup misses when the real DB title contains
# characters that ``safe_filename`` lossily replaces (``:``, ``/``, ``*``,
# etc.) — common for connector-imported docs (Google Calendar/Drive etc.).
# The workspace tree shows the lossy filename, so the agent passes that
# filename back here. Scan all documents in the resolved folder and match
# by ``safe_filename(title)`` to recover the original document.
folder_scan = select(Document).where(
Document.search_space_id == search_space_id,
)
if folder_id is None:
folder_scan = folder_scan.where(Document.folder_id.is_(None))
else:
folder_scan = folder_scan.where(Document.folder_id == folder_id)
result = await session.execute(folder_scan)
for candidate_doc in result.scalars().all():
encoded = safe_filename(str(candidate_doc.title or "untitled"))
if encoded == basename:
return candidate_doc
return None
async def _resolve_folder_id(
session: AsyncSession,
*,
search_space_id: int,
folder_parts: list[str],
) -> int | None:
"""Look up the leaf folder id for a chain of folder names; return ``None`` if missing."""
if not folder_parts:
return None
parent_id: int | None = None
for raw in folder_parts:
name = safe_folder_segment(raw)
query = select(Folder.id).where(
Folder.search_space_id == search_space_id,
Folder.name == name,
)
if parent_id is None:
query = query.where(Folder.parent_id.is_(None))
else:
query = query.where(Folder.parent_id == parent_id)
result = await session.execute(query)
row = result.first()
if row is None:
return None
parent_id = row[0]
return parent_id
def parse_documents_path(virtual_path: str) -> tuple[list[str], str]:
"""Parse a ``/documents/...`` path into ``(folder_parts, document_title)``.
The title has any ``.xml`` extension and trailing ``" (<doc_id>)"``
disambiguation suffix stripped.
"""
if not virtual_path or not virtual_path.startswith(DOCUMENTS_ROOT):
return [], ""
rel = virtual_path[len(DOCUMENTS_ROOT) :].strip("/")
if not rel:
return [], ""
parts = [p for p in rel.split("/") if p]
if not parts:
return [], ""
folder_parts = parts[:-1]
basename = parts[-1]
stem, _ = parse_doc_id_suffix(basename)
title = stem
if title.endswith(".xml"):
title = title[:-4]
return folder_parts, title
__all__ = [
"DOCUMENTS_ROOT",
"PathIndex",
"build_path_index",
"doc_to_virtual_path",
"parse_doc_id_suffix",
"parse_documents_path",
"safe_filename",
"safe_folder_segment",
"virtual_path_to_doc",
]

239
surfsense_backend/app/agents/chat/runtime/prompt_caching.py Normal file
View file

@ -0,0 +1,239 @@
r"""LiteLLM-native prompt caching configuration for SurfSense agents.
Replaces the legacy ``AnthropicPromptCachingMiddleware`` (which never
activated for our LiteLLM-based stack — its ``isinstance(model, ChatAnthropic)``
gate always failed) with LiteLLM's universal caching mechanism.
Coverage:
- Marker-based providers (need ``cache_control`` injection, which LiteLLM
performs automatically when ``cache_control_injection_points`` is set):
``anthropic/``, ``bedrock/``, ``vertex_ai/``, ``gemini/``, ``azure_ai/``,
``openrouter/`` (Claude/Gemini/MiniMax/GLM/z-ai routes), ``databricks/``
(Claude), ``dashscope/`` (Qwen), ``minimax/``, ``zai/`` (GLM).
- Auto-cached (LiteLLM strips the marker silently): ``openai/``,
``deepseek/``, ``xai/`` — these caches automatically for prompts ≥1024
tokens and surface ``prompt_cache_key`` / ``prompt_cache_retention``.
We inject **two** breakpoints per request:
- ``index: 0`` — pins the SurfSense system prompt at the head of the
request (provider variant, citation rules, tool catalog, KB tree,
skills metadata). The langchain agent factory always prepends
``request.system_message`` at index 0 (see ``factory.py``
``_execute_model_async``), so this targets exactly the main system
prompt regardless of how many other ``SystemMessage``\ s the
``before_agent`` injectors (priority, tree, memory, file-intent,
anonymous-doc) have inserted into ``state["messages"]``. Using
``role: system`` here would apply ``cache_control`` to **every**
system-role message and trip Anthropic's hard cap of 4 cache
breakpoints per request once the conversation accumulates enough
injected system messages — which surfaces as the upstream 400
``A maximum of 4 blocks with cache_control may be provided. Found N``
via OpenRouter→Anthropic.
- ``index: -1`` — pins the latest message so multi-turn savings compound:
Anthropic-family providers use longest-matching-prefix lookup, so turn
N+1 still reads turn N's cache up to the shared prefix.
For OpenAI-family configs we additionally pass:
- ``prompt_cache_key=f"surfsense-thread-{thread_id}"`` — routing hint that
raises hit rate by sending requests with a shared prefix to the same
backend. Supported by ``openai/``, ``deepseek/``, ``xai/``, and
``azure/`` (added to LiteLLM's Azure transformer in
https://github.com/BerriAI/litellm/pull/20989, Feb 2026; verified
against ``AzureOpenAIConfig.get_supported_openai_params`` in our
installed litellm 1.83.14 for ``azure/gpt-4o``, ``azure/gpt-4o-mini``,
``azure/gpt-5.4``, ``azure/gpt-5.4-mini``).
- ``prompt_cache_retention="24h"`` — extends cache TTL beyond the default
5-10 min in-memory cache. Set ONLY for OpenAI/DeepSeek/xAI: Azure's
server-side support landed in Microsoft's docs on 2026-05-13 but
LiteLLM 1.83.14's Azure transformer still omits it from its supported
params list, so it gets silently dropped by ``litellm.drop_params``.
Azure's default in-memory retention (5-10 min, max 1 h) already
bridges intra-conversation turns; revisit when LiteLLM bumps Azure.
Safety net: ``litellm.drop_params=True`` is set globally in
``app.services.llm_service`` at module-load time. Any kwarg the destination
provider doesn't recognise is auto-stripped at the provider transformer
layer, so an OpenAI→Bedrock auto-mode fallback can't 400 on
``prompt_cache_key`` etc.
"""
from __future__ import annotations
import logging
from typing import TYPE_CHECKING, Any
from langchain_core.language_models import BaseChatModel
if TYPE_CHECKING:
from app.agents.chat.runtime.llm_config import AgentConfig
logger = logging.getLogger(__name__)
# Two-breakpoint policy: head-of-request + latest message. See module
# docstring for rationale. Anthropic caps requests at 4 ``cache_control``
# blocks; we use 2 here, leaving headroom for Phase-2 tool caching.
#
# IMPORTANT: ``index: 0`` (not ``role: system``). The deepagent stack's
# ``before_agent`` middlewares (priority, tree, memory, anonymous-doc)
# insert ``SystemMessage`` instances into ``state["messages"]`` that
# accumulate across turns. With ``role: system`` the LiteLLM hook would
# tag *every* one of them with ``cache_control`` and overflow Anthropic's
# 4-block limit. ``index: 0`` always targets the langchain-prepended
# ``request.system_message``, giving us exactly one stable cache breakpoint.
_DEFAULT_INJECTION_POINTS: tuple[dict[str, Any], ...] = (
{"location": "message", "index": 0},
{"location": "message", "index": -1},
)
# Providers (uppercase ``AgentConfig.provider`` values) that accept the
# OpenAI ``prompt_cache_key`` routing hint. Microsoft's Azure OpenAI docs
# (2026-05-13) confirm automatic prompt caching applies to every GPT-4o
# or newer Azure deployment at ≥1024 tokens with no configuration needed,
# and that ``prompt_cache_key`` is combined with the prefix hash to
# improve routing affinity and therefore cache hit rate. LiteLLM's Azure
# transformer ships ``prompt_cache_key`` in its supported params as of
# https://github.com/BerriAI/litellm/pull/20989.
#
# Strict whitelist — many other providers in ``PROVIDER_MAP`` route
# through litellm's ``openai`` prefix without implementing the OpenAI
# prompt-cache surface (e.g. MOONSHOT, ZHIPU, MINIMAX), so we can't infer
# family from the litellm prefix alone.
_PROMPT_CACHE_KEY_PROVIDERS: frozenset[str] = frozenset(
{"OPENAI", "DEEPSEEK", "XAI", "AZURE", "AZURE_OPENAI"}
)
# Subset of ``_PROMPT_CACHE_KEY_PROVIDERS`` that also accept
# ``prompt_cache_retention="24h"``. Azure is excluded: see module
# docstring — LiteLLM 1.83.14's Azure transformer omits the param so
# ``drop_params`` silently strips it. Re-add Azure once a future LiteLLM
# release wires it into ``AzureOpenAIConfig.get_supported_openai_params``.
_PROMPT_CACHE_RETENTION_PROVIDERS: frozenset[str] = frozenset(
{"OPENAI", "DEEPSEEK", "XAI"}
)
def _is_router_llm(llm: BaseChatModel) -> bool:
"""Detect ``ChatLiteLLMRouter`` (auto-mode) without an eager import.
Importing ``app.services.llm_router_service`` at module-load time would
create a cycle via ``llm_config -> prompt_caching -> llm_router_service``.
Class-name comparison is sufficient since the class is defined in a
single place.
"""
return type(llm).__name__ == "ChatLiteLLMRouter"
def _provider_supports_prompt_cache_key(agent_config: AgentConfig | None) -> bool:
"""Whether the config targets a provider that accepts ``prompt_cache_key``.
Strict — only returns True for explicitly chosen OPENAI, DEEPSEEK,
XAI, AZURE, or AZURE_OPENAI providers. Auto-mode and custom
providers return False because we can't statically know the
destination and the router fans out across mixed providers.
"""
if agent_config is None or not agent_config.provider:
return False
if agent_config.is_auto_mode:
return False
if agent_config.custom_provider:
return False
return agent_config.provider.upper() in _PROMPT_CACHE_KEY_PROVIDERS
def _provider_supports_prompt_cache_retention(
agent_config: AgentConfig | None,
) -> bool:
"""Whether the config targets a provider that accepts ``prompt_cache_retention``.
Tighter than :func:`_provider_supports_prompt_cache_key` — Azure
deployments are excluded until LiteLLM ships the param in its Azure
transformer (see module docstring).
"""
if agent_config is None or not agent_config.provider:
return False
if agent_config.is_auto_mode:
return False
if agent_config.custom_provider:
return False
return agent_config.provider.upper() in _PROMPT_CACHE_RETENTION_PROVIDERS
def _get_or_init_model_kwargs(llm: BaseChatModel) -> dict[str, Any] | None:
"""Return ``llm.model_kwargs`` as a writable dict, or ``None`` to bail.
Initialises the field to ``{}`` when present-but-None on a Pydantic v2
model. Returns ``None`` if the LLM type doesn't expose a writable
``model_kwargs`` attribute (caller should treat as no-op).
"""
model_kwargs = getattr(llm, "model_kwargs", None)
if isinstance(model_kwargs, dict):
return model_kwargs
try:
llm.model_kwargs = {} # type: ignore[attr-defined]
except Exception:
return None
refreshed = getattr(llm, "model_kwargs", None)
return refreshed if isinstance(refreshed, dict) else None
def apply_litellm_prompt_caching(
llm: BaseChatModel,
*,
agent_config: AgentConfig | None = None,
thread_id: int | None = None,
) -> None:
"""Configure LiteLLM prompt caching on a ChatLiteLLM/ChatLiteLLMRouter.
Idempotent — values already present in ``llm.model_kwargs`` (e.g. from
``agent_config.litellm_params`` overrides) are preserved. Mutates
``llm.model_kwargs`` in place; the kwargs flow to ``litellm.completion``
via ``ChatLiteLLM._default_params`` and via ``self.model_kwargs`` merge
in our custom ``ChatLiteLLMRouter``.
Args:
llm: ChatLiteLLM, SanitizedChatLiteLLM, or ChatLiteLLMRouter instance.
agent_config: Optional ``AgentConfig`` driving provider-specific
behaviour. When omitted (or auto-mode), only the universal
``cache_control_injection_points`` are set.
thread_id: Optional thread id used to construct a per-thread
``prompt_cache_key`` for OpenAI-family providers. Caching still
works without it (server-side automatic), but the key improves
backend routing affinity and therefore hit rate.
"""
model_kwargs = _get_or_init_model_kwargs(llm)
if model_kwargs is None:
logger.debug(
"apply_litellm_prompt_caching: %s exposes no writable model_kwargs; skipping",
type(llm).__name__,
)
return
if "cache_control_injection_points" not in model_kwargs:
model_kwargs["cache_control_injection_points"] = [
dict(point) for point in _DEFAULT_INJECTION_POINTS
]
# OpenAI-style extras only when we statically know the destination
# accepts them. Auto-mode router fans out across mixed providers so
# we can't safely set destination-specific kwargs there (drop_params
# would strip them but it's wasteful to set them in the first
# place).
if _is_router_llm(llm):
return
if (
thread_id is not None
and "prompt_cache_key" not in model_kwargs
and _provider_supports_prompt_cache_key(agent_config)
):
model_kwargs["prompt_cache_key"] = f"surfsense-thread-{thread_id}"
if (
"prompt_cache_retention" not in model_kwargs
and _provider_supports_prompt_cache_retention(agent_config)
):
model_kwargs["prompt_cache_retention"] = "24h"