langfuse and conversiotn memory trimmed
Browse files
config.py
CHANGED
|
@@ -17,6 +17,9 @@ DEFAULT_MAX_TOKENS = 2048
|
|
| 17 |
DEFAULT_DATA_DIR = "data"
|
| 18 |
DEFAULT_LANGFUSE_HOST = "https://cloud.langfuse.com"
|
| 19 |
DEFAULT_CHECKPOINT_DB = "checkpoints.sqlite"
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
|
| 22 |
class ConfigError(RuntimeError):
|
|
@@ -31,6 +34,7 @@ class Settings:
|
|
| 31 |
model: str
|
| 32 |
temperature: float
|
| 33 |
max_tokens: int
|
|
|
|
| 34 |
checkpoint_db: Path
|
| 35 |
langfuse_public_key: str
|
| 36 |
langfuse_secret_key: str
|
|
@@ -95,6 +99,7 @@ def get_settings() -> Settings:
|
|
| 95 |
model=os.getenv("GROQ_MODEL", "").strip() or DEFAULT_MODEL,
|
| 96 |
temperature=_get_float("LLM_TEMPERATURE", DEFAULT_TEMPERATURE),
|
| 97 |
max_tokens=_get_int("LLM_MAX_TOKENS", DEFAULT_MAX_TOKENS),
|
|
|
|
| 98 |
checkpoint_db=PROJECT_ROOT / (
|
| 99 |
os.getenv("CHECKPOINT_DB", "").strip() or DEFAULT_CHECKPOINT_DB
|
| 100 |
),
|
|
|
|
| 17 |
DEFAULT_DATA_DIR = "data"
|
| 18 |
DEFAULT_LANGFUSE_HOST = "https://cloud.langfuse.com"
|
| 19 |
DEFAULT_CHECKPOINT_DB = "checkpoints.sqlite"
|
| 20 |
+
# Turns of conversation replayed to the model. Older turns stay in the
|
| 21 |
+
# checkpoint (nothing is deleted) but are not resent.
|
| 22 |
+
DEFAULT_MAX_HISTORY_TURNS = 10
|
| 23 |
|
| 24 |
|
| 25 |
class ConfigError(RuntimeError):
|
|
|
|
| 34 |
model: str
|
| 35 |
temperature: float
|
| 36 |
max_tokens: int
|
| 37 |
+
max_history_turns: int
|
| 38 |
checkpoint_db: Path
|
| 39 |
langfuse_public_key: str
|
| 40 |
langfuse_secret_key: str
|
|
|
|
| 99 |
model=os.getenv("GROQ_MODEL", "").strip() or DEFAULT_MODEL,
|
| 100 |
temperature=_get_float("LLM_TEMPERATURE", DEFAULT_TEMPERATURE),
|
| 101 |
max_tokens=_get_int("LLM_MAX_TOKENS", DEFAULT_MAX_TOKENS),
|
| 102 |
+
max_history_turns=_get_int("MAX_HISTORY_TURNS", DEFAULT_MAX_HISTORY_TURNS),
|
| 103 |
checkpoint_db=PROJECT_ROOT / (
|
| 104 |
os.getenv("CHECKPOINT_DB", "").strip() or DEFAULT_CHECKPOINT_DB
|
| 105 |
),
|
graph.py
CHANGED
|
@@ -9,6 +9,7 @@ rewrite.
|
|
| 9 |
import sqlite3
|
| 10 |
from typing import Annotated, TypedDict
|
| 11 |
|
|
|
|
| 12 |
from langchain_groq import ChatGroq
|
| 13 |
from langgraph.checkpoint.sqlite import SqliteSaver
|
| 14 |
from langgraph.graph import START, StateGraph
|
|
@@ -40,8 +41,20 @@ def build_graph():
|
|
| 40 |
profile_data = ProfileLoader(settings.data_dir).load_and_format_markdown()
|
| 41 |
chain = build_qa_prompt() | build_llm()
|
| 42 |
|
|
|
|
|
|
|
|
|
|
| 43 |
def answer(state: ChatState):
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
user_query = state["messages"][-1].content
|
| 46 |
response = chain.invoke({
|
| 47 |
"context_text": profile_data,
|
|
|
|
| 9 |
import sqlite3
|
| 10 |
from typing import Annotated, TypedDict
|
| 11 |
|
| 12 |
+
from langchain_core.messages import trim_messages
|
| 13 |
from langchain_groq import ChatGroq
|
| 14 |
from langgraph.checkpoint.sqlite import SqliteSaver
|
| 15 |
from langgraph.graph import START, StateGraph
|
|
|
|
| 41 |
profile_data = ProfileLoader(settings.data_dir).load_and_format_markdown()
|
| 42 |
chain = build_qa_prompt() | build_llm()
|
| 43 |
|
| 44 |
+
# Two messages per turn (human + ai).
|
| 45 |
+
max_history_messages = settings.max_history_turns * 2
|
| 46 |
+
|
| 47 |
def answer(state: ChatState):
|
| 48 |
+
# Older turns remain in the checkpoint; they are simply not resent,
|
| 49 |
+
# which keeps per-turn cost flat instead of growing with the thread.
|
| 50 |
+
history = trim_messages(
|
| 51 |
+
state["messages"][:-1],
|
| 52 |
+
max_tokens=max_history_messages,
|
| 53 |
+
token_counter=len,
|
| 54 |
+
strategy="last",
|
| 55 |
+
start_on="human",
|
| 56 |
+
include_system=False,
|
| 57 |
+
)
|
| 58 |
user_query = state["messages"][-1].content
|
| 59 |
response = chain.invoke({
|
| 60 |
"context_text": profile_data,
|