mirror of
https://github.com/imartinez/privateGPT.git
synced 2026-08-08 23:35:29 +00:00
* feat: add bundle to remove * fix: spaces * feat: add xml render as skill spec * feat: add skill volume root to cache * fix: deduplicate values * fix: change the mount path to skill_id * fix: use different paths * feat: add principal (cherry picked from commit 5db64fe721d5706440ce9f342b1388ffe742bc16) # Conflicts: # private_gpt/components/code_execution/base.py # private_gpt/components/code_execution/code_execution_component.py # private_gpt/components/code_execution/local.py # private_gpt/components/environment/manager.py # private_gpt/components/tools/builders/bash_tool_builder.py # private_gpt/components/tools/builders/text_editor_tool_builder.py # private_gpt/components/tools/processors/bash_processor.py * fix: mypy ... * fix: config * fix: sandbox config * feat: add forward cookies * fix: loop * feat: add present server * feat: add feature flag for tools * refactor: move principal to another better place * feat: add api key principal * docs: fix docs * docs: add present server * fix: principal * fix: mypy * fix: avoid to block the loop * fix: blocks in expansion * fix: remove maximum concurrent users ... * fix: multiplexer * fix: readers * fix: more fixes ... * fix: impl * feat: tool scheduler * feat: add adaptative * feat: add chat worker * fix: config * fix: max * feat: add chat/tools workers * fix: mypy * feat: add generic scheduler * fix: get result * feat: do serializable the tool executor * fix: tools * fix: config * fix: config * fix: args * fix: config * fix: serializer * Revert "fix: blocks in expansion" This reverts commita2110f94a8. * fix: unify all logic * feat: add ingestion scheduler * fix: settings * fix: config * feat: add arq worker to chat * fix: arq worker * fix: add nest * fix: mypy * fix: await * fix: script stress * fix: tokenizer * fix: chat scheduler * fix: mypy * fix: add async tokenizer * fix: improve condense * fix: tool scheduler * feat: add initial real async chat worker * fix: mypy * fix: do resumable local executor ... ... ... fix: revert usleess changes fix: remove parent chat job fix: refactor fix: loop ref: rename models fix: chat engine fix: mypy ... ... ... fix: fix deps * fix: tests * fix: tests * ... * fix: stream * fix: config * fix: scheduler * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * Handle PGPT_WORKER_MODE=celery in health check worker status * fix: cancel * fix: arch * fix: test ingestion * fix: deserialization of chat messages * fix: broken results * fix: mypy * fix: test * fix: config * fix: remove arq tool worker * fix: output cls * fix: preserve early resumable tool callbacks * fix: preserve async tool result order * refactor: address worker PR review comments * fix: mypy * test: colocate ARQ chat enqueue coverage * fix: remove redis from tests * fix: mypy * fix: tests * test: isolate chat mocks and cancellation timing * fix: tests * fix: tests * fix: test (cherry picked from commitf8ee460af2) * fix: worker config * test: remove flaky chat cancellation assertion (cherry picked from commit1115ff2349) # Conflicts: # tests/server/chat/test_chat_routes.py * fix: emit chat pings from stream listeners * fix: don't duplciate the output * fix: rss memory ... * fix: pass the args * fix: threads * fix: websearch --------- Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
251 lines
8.5 KiB
Python
251 lines
8.5 KiB
Python
from typing import Annotated, Literal
|
|
|
|
from llama_index.core.base.llms.types import TextBlock
|
|
from pydantic import BaseModel, ConfigDict, Field
|
|
|
|
from private_gpt.components.chat.models.chat_config_models import ToolSpec
|
|
from private_gpt.components.context.models.layer_type import LayerType
|
|
from private_gpt.components.engines.citations.types import Document
|
|
from private_gpt.components.sandbox.content_bundle import ContentBundle
|
|
|
|
|
|
class BaseContextLayer(BaseModel):
|
|
"""Common fields shared by every context layer."""
|
|
|
|
source: str = Field(
|
|
default="request",
|
|
description="Origin of the layer, e.g. 'platform', 'skill:git', 'mcp'.",
|
|
)
|
|
priority: int = Field(
|
|
default=1000,
|
|
description=(
|
|
"Render priority for prompt layers. Lower values are rendered first. "
|
|
"State-only layers ignore this field."
|
|
),
|
|
)
|
|
|
|
model_config = ConfigDict(frozen=True, arbitrary_types_allowed=True)
|
|
|
|
def render(self) -> str:
|
|
"""Return text to include in the system prompt (empty for state layers)."""
|
|
return ""
|
|
|
|
|
|
class UserInstructionsLayer(BaseContextLayer):
|
|
"""User/system provided baseline instructions."""
|
|
|
|
type: Literal[LayerType.USER_INSTRUCTIONS] = Field(
|
|
default=LayerType.USER_INSTRUCTIONS, frozen=True
|
|
)
|
|
priority: int = Field(default=100, frozen=True)
|
|
text: str | list[TextBlock] | None = Field(description="Raw instruction text.")
|
|
|
|
def render(self) -> str:
|
|
if self.text is None:
|
|
return ""
|
|
if isinstance(self.text, str):
|
|
return self.text.strip()
|
|
|
|
texts = [block.text.strip() for block in self.text if block.text.strip()]
|
|
return "\n\n".join(texts)
|
|
|
|
|
|
class RuntimeInstructionsLayer(BaseContextLayer):
|
|
"""Transient runtime instructions (e.g. condensation hints)."""
|
|
|
|
type: Literal[LayerType.RUNTIME_INSTRUCTIONS] = Field(
|
|
default=LayerType.RUNTIME_INSTRUCTIONS, frozen=True
|
|
)
|
|
priority: int = Field(default=200, frozen=True)
|
|
text: str = Field(description="Additional instruction text.")
|
|
|
|
def render(self) -> str:
|
|
return self.text
|
|
|
|
|
|
class SkillCatalogEntry(BaseModel):
|
|
id: str = Field(description="Skill identifier.")
|
|
name: str = Field(description="Skill frontmatter name.")
|
|
description: str = Field(description="Skill frontmatter description.")
|
|
loading: Literal["eager", "lazy"] = Field(description="Skill loading mode.")
|
|
location: str = Field(
|
|
default="",
|
|
description="Path to the skill's SKILL.md inside the execution "
|
|
"environment, e.g. /mnt/skills/pdf/SKILL.md.",
|
|
)
|
|
resources: list[str] = Field(
|
|
default_factory=list,
|
|
description="Bundled file paths relative to the skill directory.",
|
|
)
|
|
|
|
|
|
class SkillCatalogLayer(BaseContextLayer):
|
|
"""Catalog of available-but-not-yet-loaded skills shown to the LLM."""
|
|
|
|
type: Literal[LayerType.SKILL_CATALOG] = Field(
|
|
default=LayerType.SKILL_CATALOG, frozen=True
|
|
)
|
|
priority: int = Field(default=300, frozen=True)
|
|
entries: list[SkillCatalogEntry] = Field(
|
|
default_factory=list,
|
|
description="List of available skill entries.",
|
|
)
|
|
|
|
def render(self) -> str:
|
|
if not self.entries:
|
|
return ""
|
|
lines = ["<available_skills>"]
|
|
for entry in self.entries:
|
|
lines.append(" <skill>")
|
|
lines.append(f" <name>{entry.name}</name>")
|
|
lines.append(f" <description>{entry.description}</description>")
|
|
if entry.location:
|
|
lines.append(f" <location>{entry.location}</location>")
|
|
if entry.resources:
|
|
lines.append(" <resources>")
|
|
lines.extend(
|
|
f" <resource>{resource}</resource>"
|
|
for resource in entry.resources
|
|
)
|
|
lines.append(" </resources>")
|
|
lines.append(" </skill>")
|
|
lines.append("</available_skills>")
|
|
return "\n".join(lines)
|
|
|
|
|
|
class SkillBodyLayer(BaseContextLayer):
|
|
"""Full instructions for one activated skill."""
|
|
|
|
type: Literal[LayerType.SKILL_BODY] = Field(
|
|
default=LayerType.SKILL_BODY, frozen=True
|
|
)
|
|
priority: int = Field(default=400, frozen=True)
|
|
skill_id: str = Field(description="Skill identifier.")
|
|
name: str = Field(description="Skill frontmatter name.")
|
|
version: str = Field(description="Skill version token.")
|
|
instructions: str = Field(description="Skill instruction body content.")
|
|
location: str = Field(
|
|
default="",
|
|
description="Skill directory inside the execution environment, "
|
|
"e.g. /mnt/skills/pdf/.",
|
|
)
|
|
resources: list[str] = Field(
|
|
default_factory=list,
|
|
description="Bundled file paths relative to the skill directory.",
|
|
)
|
|
render_as_xml: bool = Field(
|
|
default=True,
|
|
description="When True, wrap output in <skill_content> XML tags. "
|
|
"Eager skills render as plain text (False).",
|
|
)
|
|
|
|
def render(self) -> str:
|
|
body = self.instructions.strip()
|
|
parts: list[str] = []
|
|
|
|
if self.location:
|
|
parts.append(f"Skill directory: {self.location}")
|
|
parts.append(
|
|
"Relative paths in this skill are relative to the skill directory."
|
|
)
|
|
|
|
if self.resources:
|
|
resource_lines = ["<skill_resources>"]
|
|
resource_lines.extend(f" <file>{r}</file>" for r in self.resources)
|
|
resource_lines.append("</skill_resources>")
|
|
parts.append("\n".join(resource_lines))
|
|
|
|
footer = "\n\n".join(parts)
|
|
inner = f"{body}\n\n{footer}" if body and footer else (body or footer)
|
|
|
|
if self.render_as_xml:
|
|
return f'<skill_content name="{self.name}">\n{inner}\n</skill_content>'
|
|
return inner
|
|
|
|
|
|
class ToolInstructionsLayer(BaseContextLayer):
|
|
"""Per-tool instructions injected when a tool is available."""
|
|
|
|
type: Literal[LayerType.TOOL_INSTRUCTIONS] = Field(
|
|
default=LayerType.TOOL_INSTRUCTIONS, frozen=True
|
|
)
|
|
priority: int = Field(default=450, frozen=True)
|
|
tool_name: str = Field(
|
|
description="Canonical tool name these instructions apply to."
|
|
)
|
|
instructions: str = Field(description="Instruction text for this tool.")
|
|
|
|
def render(self) -> str:
|
|
return self.instructions
|
|
|
|
|
|
class DocumentLayer(BaseContextLayer):
|
|
"""One document injected as context — wraps the real Document entity."""
|
|
|
|
type: Literal[LayerType.DOCUMENT] = Field(default=LayerType.DOCUMENT, frozen=True)
|
|
priority: int = Field(default=2000, frozen=True)
|
|
document: Document = Field(description="The Document domain object.")
|
|
|
|
def render(self) -> str:
|
|
return "" # state layer — never in system prompt
|
|
|
|
|
|
class ToolDefinitionsLayer(BaseContextLayer):
|
|
"""Tool specs available to the LLM — consumed programmatically, not rendered."""
|
|
|
|
type: Literal[LayerType.TOOL_DEFINITIONS] = Field(
|
|
default=LayerType.TOOL_DEFINITIONS, frozen=True
|
|
)
|
|
priority: int = Field(default=2000, frozen=True)
|
|
tools: list[ToolSpec] = Field(
|
|
default_factory=list,
|
|
description="List of ToolSpec instances.",
|
|
)
|
|
|
|
def render(self) -> str:
|
|
return "" # state layer — never in system prompt
|
|
|
|
|
|
class ContextPromptLayer(BaseContextLayer):
|
|
"""Rendered context section to include in the system prompt."""
|
|
|
|
type: Literal[LayerType.CONTEXT] = Field(default=LayerType.CONTEXT, frozen=True)
|
|
priority: int = Field(default=600, frozen=True)
|
|
text: str = Field(description="Rendered context prompt text.")
|
|
|
|
def render(self) -> str:
|
|
return self.text
|
|
|
|
|
|
class ContentBundlesLayer(BaseContextLayer):
|
|
"""Skill content bundles — consumed by tool builders, not rendered."""
|
|
|
|
type: Literal[LayerType.CONTENT_BUNDLES] = Field(
|
|
default=LayerType.CONTENT_BUNDLES, frozen=True
|
|
)
|
|
priority: int = Field(default=2000, frozen=True)
|
|
bundles: list[ContentBundle] = Field(default_factory=list)
|
|
to_remove: list[str] = Field(
|
|
default_factory=list,
|
|
description="Canonical paths of skill bundles to remove from the sandbox.",
|
|
)
|
|
|
|
model_config = ConfigDict(frozen=True, arbitrary_types_allowed=True)
|
|
|
|
def render(self) -> str:
|
|
return "" # state layer — never in system prompt
|
|
|
|
|
|
AnyContextLayer = Annotated[
|
|
UserInstructionsLayer
|
|
| RuntimeInstructionsLayer
|
|
| ContextPromptLayer
|
|
| SkillCatalogLayer
|
|
| SkillBodyLayer
|
|
| ToolInstructionsLayer
|
|
| DocumentLayer
|
|
| ToolDefinitionsLayer
|
|
| ContentBundlesLayer,
|
|
Field(discriminator="type"),
|
|
]
|