feat(agents-api): Adaptive context (via recursive summarization)

agents-api/agents_api/activities/logger.py

@@ -0,0 +1,8 @@
+import logging
+
+
+logger = logging.getLogger(__name__)
+h = logging.StreamHandler()
+fmt = logging.Formatter("[%(asctime)s/%(levelname)s] - %(message)s")
+h.setFormatter(fmt)
+logger.addHandler(h)

agents-api/agents_api/activities/summarization.py

@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 
+import asyncio
 from uuid import UUID
 from typing import Callable
 from textwrap import dedent
@@ -11,7 +12,10 @@
 )
 from agents_api.common.protocol.entries import Entry
 from ..model_registry import JULEP_MODELS
-from ..env import summarization_model_name, model_inference_url, model_api_key
+from ..env import model_inference_url, model_api_key, summarization_model_name
+from agents_api.rec_sum.entities import get_entities
+from agents_api.rec_sum.summarize import summarize_messages
+from agents_api.rec_sum.trim import trim_messages
 
 
 example_previous_memory = """
@@ -160,28 +164,51 @@ async def run_prompt(
 @activity.defn
 async def summarization(session_id: str) -> None:
     session_id = UUID(session_id)
-    entries = [
-        Entry(**row)
-        for _, row in get_toplevel_entries_query(session_id=session_id).iterrows()
-    ]
+    entries = []
+    entities_entry_ids = []
+    for _, row in get_toplevel_entries_query(session_id=session_id).iterrows():
+        if row["role"] == "system" and row.get("name") == "entities":
+            entities_entry_ids.append(UUID(row["entry_id"], version=4))
+        else:
+            entries.append(row)
 
     assert len(entries) > 0, "no need to summarize on empty entries list"
 
-    response = await run_prompt(
-        dialog=entries, previous_memories=[], model=f"openai/{summarization_model_name}"
+    trimmed_messages, entities = await asyncio.gather(
+        trim_messages(entries, model=summarization_model_name),
+        get_entities(entries, model=summarization_model_name),
     )
-
-    new_entry = Entry(
+    summarized = await summarize_messages(trimmed_messages)
+    ts_delta = (entries[1]["timestamp"] - entries[0]["timestamp"]) / 2
+    new_entities_entry = Entry(
         session_id=session_id,
         source="summarizer",
         role="system",
-        name="information",
-        content=response,
-        timestamp=entries[-1].timestamp + 0.01,
+        name="entities",
+        content=entities["content"],
+        timestamp=entries[0]["timestamp"] + ts_delta,
     )
 
     entries_summarization_query(
         session_id=session_id,
-        new_entry=new_entry,
-        old_entry_ids=[e.id for e in entries],
+        new_entry=new_entities_entry,
+        old_entry_ids=entities_entry_ids,
     )
+
+    for msg in summarized:
+        new_entry = Entry(
+            session_id=session_id,
+            source="summarizer",
+            role="system",
+            name="information",
+            content=msg["content"],
+            timestamp=entries[-1]["timestamp"] + 0.01,
+        )
+
+        entries_summarization_query(
+            session_id=session_id,
+            new_entry=new_entry,
+            old_entry_ids=[
+                UUID(entries[idx]["entry_id"], version=4) for idx in msg["summarizes"]
+            ],
+        )

agents-api/agents_api/activities/truncation.py

@@ -0,0 +1,51 @@
+from uuid import UUID
+from temporalio import activity
+from agents_api.models.entry.entries_summarization import get_toplevel_entries_query
+from agents_api.models.entry.delete_entries import delete_entries
+from agents_api.autogen.openapi_model import Role
+from agents_api.common.protocol.entries import Entry
+
+
+def get_extra_entries(messages: list[Entry], token_count_threshold: int) -> list[Entry]:
+    if not len(messages):
+        return messages
+
+    result: list[Entry] = []
+    token_cnt, offset = 0, 0
+    if messages[0].role == Role.system:
+        token_cnt, offset = messages[0].token_count, 1
+
+    for m in reversed(messages[offset:]):
+        token_cnt += m.token_count
+        if token_cnt < token_count_threshold:
+            continue
+        else:
+            result.append(m)
+
+    return result
+
+
+@activity.defn
+async def truncation(session_id: str, token_count_threshold: int) -> None:
+    session_id = UUID(session_id)
+
+    delete_entries(
+        get_extra_entries(
+            [
+                Entry(
+                    entry_id=row["entry_id"],
+                    session_id=session_id,
+                    source=row["source"],
+                    role=Role(row["role"]),
+                    name=row["name"],
+                    content=row["content"],
+                    created_at=row["created_at"],
+                    timestamp=row["timestamp"],
+                )
+                for _, row in get_toplevel_entries_query(
+                    session_id=session_id
+                ).iterrows()
+            ],
+            token_count_threshold,
+        ),
+    )

agents-api/agents_api/autogen/openapi_model.py

@@ -1,6 +1,6 @@
 # generated by datamodel-codegen:
 #   filename:  openapi.yaml
-#   timestamp: 2024-05-28T03:07:50+00:00
+#   timestamp: 2024-06-03T12:10:34+00:00
 
 from __future__ import annotations
 
@@ -124,6 +124,14 @@ class Session(BaseModel):
     """
     Render system and assistant message content as jinja templates
     """
+    token_budget: int | None = None
+    """
+    Threshold value for the adaptive context functionality
+    """
+    context_overflow: str | None = None
+    """
+    Action to start on context window overflow
+    """
 
 
 class CreateSessionRequest(BaseModel):
@@ -151,6 +159,14 @@ class CreateSessionRequest(BaseModel):
     """
     Render system and assistant message content as jinja templates
     """
+    token_budget: int | None = None
+    """
+    Threshold value for the adaptive context functionality
+    """
+    context_overflow: str | None = None
+    """
+    Action to start on context window overflow
+    """
 
 
 class UpdateSessionRequest(BaseModel):
@@ -166,6 +182,14 @@ class UpdateSessionRequest(BaseModel):
     """
     Optional metadata
     """
+    token_budget: int | None = None
+    """
+    Threshold value for the adaptive context functionality
+    """
+    context_overflow: str | None = None
+    """
+    Action to start on context window overflow
+    """
 
 
 class UpdateUserRequest(BaseModel):
@@ -753,6 +777,14 @@ class PatchSessionRequest(BaseModel):
     """
     Optional metadata
     """
+    token_budget: int | None = None
+    """
+    Threshold value for the adaptive context functionality
+    """
+    context_overflow: str | None = None
+    """
+    Action to start on context window overflow
+    """
 
 
 class PartialFunctionDef(BaseModel):

agents-api/agents_api/clients/temporal.py

@@ -46,3 +46,16 @@ async def run_embed_docs_task(
         task_queue="memory-task-queue",
         id=str(job_id),
     )
+
+
+async def run_truncation_task(
+    token_count_threshold: int, session_id: UUID, job_id: UUID
+):
+    client = await get_client()
+
+    await client.execute_workflow(
+        "TruncationWorkflow",
+        args=[str(session_id), token_count_threshold],
+        task_queue="memory-task-queue",
+        id=str(job_id),
+    )

agents-api/agents_api/common/protocol/entries.py

@@ -7,13 +7,18 @@
     Role,
     ChatMLImageContentPart,
     ChatMLTextContentPart,
+    Detail,
 )
 from agents_api.common.utils.datetime import utcnow
 
 EntrySource = Literal["api_request", "api_response", "internal", "summarizer"]
 Tokenizer = Literal["character_count"]
 
 
+LOW_IMAGE_TOKEN_COUNT = 85
+HIGH_IMAGE_TOKEN_COUNT = 85 + 4 * 170
+
+
 class Entry(BaseModel):
     """Represents an entry in the system, encapsulating all necessary details such as ID, session ID, source, role, and content among others."""
 
@@ -44,12 +49,15 @@ def token_count(self) -> int:
             elif isinstance(self.content, dict):
                 content_length = len(json.dumps(self.content))
             elif isinstance(self.content, list):
-                text = ""
                 for part in self.content:
-                    # TODO: how to calc token count for images?
                     if isinstance(part, ChatMLTextContentPart):
-                        text += part.text
-                content_length = len(text)
+                        content_length += len(part.text)
+                    elif isinstance(part, ChatMLImageContentPart):
+                        content_length += (
+                            LOW_IMAGE_TOKEN_COUNT
+                            if part.image_url.detail == Detail.low
+                            else HIGH_IMAGE_TOKEN_COUNT
+                        )
 
             # Divide the content length by 3.5 to estimate token count based on character count.
             return int(content_length // 3.5)

agents-api/agents_api/common/protocol/sessions.py

@@ -44,3 +44,5 @@ class SessionData(BaseModel):
     metadata: Dict = {}
     user_metadata: Optional[Dict] = None
     agent_metadata: Dict = {}
+    token_budget: int | None = None
+    context_overflow: str | None = None

agents-api/agents_api/common/utils/messages.py

@@ -1,4 +1,5 @@
 import json
+from typing import cast
 from agents_api.autogen.openapi_model import (
     ChatMLTextContentPart,
     ChatMLImageContentPart,
@@ -23,3 +24,17 @@ def content_to_json(
         result = [{"type": "text", "text": json.dumps(content, indent=4)}]
 
     return result
+
+
+def stringify_content(
+    msg: str | list[ChatMLTextContentPart] | list[ChatMLImageContentPart] | dict,
+) -> str:
+    content = ""
+    if isinstance(msg, list):
+        content = " ".join([part.text for part in msg if part.type == "text"])
+    elif isinstance(msg, str):
+        content = msg
+    elif isinstance(msg, dict) and msg["type"] == "text":
+        content = cast(str, msg["text"])
+
+    return content

agents-api/agents_api/env.py

@@ -26,12 +26,6 @@
 model_api_key: str = env.str("MODEL_API_KEY", default=None)
 model_inference_url: str = env.str("MODEL_INFERENCE_URL", default=None)
 openai_api_key: str = env.str("OPENAI_API_KEY", default="")
-summarization_ratio_threshold: float = env.float(
-    "MAX_TOKENS_RATIO_TO_SUMMARIZE", default=0.5
-)
-summarization_tokens_threshold: int = env.int(
-    "SUMMARIZATION_TOKENS_THRESHOLD", default=2048
-)
 summarization_model_name: str = env.str(
     "SUMMARIZATION_MODEL_NAME", default="gpt-4-turbo"
 )
@@ -78,8 +72,6 @@
     debug=debug,
     cozo_host=cozo_host,
     cozo_auth=cozo_auth,
-    summarization_ratio_threshold=summarization_ratio_threshold,
-    summarization_tokens_threshold=summarization_tokens_threshold,
     worker_url=worker_url,
     sentry_dsn=sentry_dsn,
     temporal_endpoint=temporal_endpoint,

agents-api/agents_api/models/entry/delete_entries.py

@@ -4,6 +4,8 @@
 
 
 from ..utils import cozo_query
+from ...common.protocol.entries import Entry
+from ...common.utils.messages import content_to_json
 
 
 @cozo_query
@@ -61,3 +63,55 @@ def delete_entries_query(session_id: UUID) -> tuple[str, dict]:
     }"""
 
     return (query, {"session_id": str(session_id)})
+
+
+@cozo_query
+def delete_entries(entries: list[Entry]) -> tuple[str, dict]:
+    entry_keys = [
+        [
+            f'to_uuid("{e.id}")',
+            f'to_uuid("{e.session_id}")',
+            e.source,
+            e.role.value if hasattr(e.role, "value") else e.role,
+            e.name,
+            content_to_json(e.content),
+            e.token_count,
+            e.tokenizer,
+            e.created_at,
+            e.timestamp,
+        ]
+        for e in entries
+    ]
+
+    query = """
+    {
+        ?[
+            entry_id,
+            session_id,
+            source,
+            role,
+            name,
+            content,
+            token_count,
+            tokenizer,
+            created_at,
+            timestamp,
+        ] <- $entry_keys
+
+        :delete entries {
+            entry_id,
+            session_id,
+            source,
+            role,
+            name,
+            content,
+            token_count,
+            tokenizer,
+            created_at,
+            timestamp,
+        }
+
+        :returning
+    }"""
+
+    return (query, {"entry_keys": entry_keys})