sanatan_ai / chat_utils.py
vikramvasudevan's picture
Upload folder using huggingface_hub
bec50ac verified
import uuid
import json
import random
import asyncio
import logging
import time
import traceback
from html import escape
from langchain_core.messages.ai import AIMessageChunk
from langchain_core.messages.system import SystemMessage
from langchain_core.messages.tool import ToolMessage
from graph_helper import generate_graph
# Logging
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
thinking_verbs = [
"thinking",
"processing",
"crunching data",
"please wait",
"just a few more seconds",
"closing in",
"analyzing",
"reasoning",
"computing",
"synthesizing insight",
"searching through the cosmos",
"decoding ancient knowledge",
"scanning the scriptures",
"accessing divine memory",
"gathering wisdom",
"consulting the rishis",
"listening to the ātmā",
"channeling sacred energy",
"unfolding the divine word",
"meditating on the meaning",
"reciting from memory",
"traversing the Vedas",
"seeking the inner light",
"invoking paramārtha",
"putting it all together",
"digging deeper",
"making sense of it",
"connecting the dots",
"almost there",
"getting closer",
"wrapping it up",
"piecing it together",
"swirling through verses",
"diving into the ocean of knowledge",
"lighting the lamp of understanding",
"walking the path of inquiry",
"aligning stars of context",
]
graph = generate_graph()
def add_node_to_tree(
node_tree: list[str], node_label: str, tooltip: str = "no arguments to show"
) -> list[str]:
if tooltip:
tooltip = escape(tooltip).replace("'", "'")
node_with_tooltip = (
f"<span class='node-label' title='{tooltip}'>{node_label}</span>"
)
else:
node_with_tooltip = node_label
node_tree[-1] = node_with_tooltip
node_tree.append("<span class='spinner'>&nbsp;</span>")
return node_tree
def end_node_tree(node_tree: list[str]) -> list[str]:
node_tree[-1] = "🏁"
return node_tree
def get_args_for_toolcall(tool_calls_buffer: dict, tool_call_id: str):
return (
tool_calls_buffer[tool_call_id]["args_str"]
if tool_call_id in tool_calls_buffer
and "args_str" in tool_calls_buffer[tool_call_id]
else ""
)
async def chat_wrapper(
message, history, thread_id, debug, preferred_language="English"
):
if debug:
async for chunk in chat_streaming(
debug, message, history, thread_id, preferred_language=preferred_language
):
yield chunk
else:
response = chat(
debug, message, history, thread_id, preferred_language=preferred_language
)
yield response
def chat(debug_mode, message, history, thread_id, preferred_language="English"):
config = {"configurable": {"thread_id": thread_id}, "recursion_limit": 30}
response = graph.invoke(
{
"debug_mode": debug_mode,
"messages": [{"role": "user", "content": message}],
"language": preferred_language,
},
config=config,
)
return response["messages"][-1].content
async def chat_streaming(
debug_mode: bool, message, history, thread_id, preferred_language="English"
):
state = {
"debug_mode": debug_mode,
"messages": (history or []) + [{"role": "user", "content": message}],
"language": preferred_language,
}
config = {"configurable": {"thread_id": thread_id}, "recursion_limit": 30}
start_time = time.time()
streamed_response = ""
final_response = ""
# final_node = "validator"
MAX_CONTENT = 500
try:
node_tree = ["🚩", "<span class='spinner'>&nbsp;</span>"]
tool_calls_buffer = {}
async for msg, metadata in graph.astream(
state, config=config, stream_mode="messages"
):
node = metadata.get("langgraph_node", "?")
name = getattr(msg, "name", "-")
if not isinstance(msg, ToolMessage):
node_icon = "🧠"
else:
node_icon = "⚙️"
node_label = f"{node}"
tool_label = f"{name or ''}"
if tool_label:
node_label = node_label + f":{tool_label}"
label = f"{node_icon} {node_label}"
tooltip = ""
if isinstance(msg, ToolMessage):
tooltip = get_args_for_toolcall(tool_calls_buffer, msg.tool_call_id)
# logger.info("tooltip = ", tooltip)
# checking for -2 last but one. since last entry is always a spinner
if node_tree[-2] != label:
add_node_to_tree(node_tree, label, tooltip)
full: str = escape(msg.content)
truncated = (full[:MAX_CONTENT] + "…") if len(full) > MAX_CONTENT else full
def generate_processing_message():
return f"<div class='thinking-bubble'><em>🤔{random.choice(thinking_verbs)} ...</em></div>"
if (
not isinstance(msg, ToolMessage)
and not isinstance(msg, SystemMessage)
and not isinstance(msg, AIMessageChunk)
):
logger.info("msg = %s", msg)
if isinstance(msg, ToolMessage):
logger.debug("tool message = %s", msg)
html = f"<div class='thinking-bubble'><em>🤔 {msg.name} tool: {random.choice(thinking_verbs)} ...</em></div>"
yield f"### { ' → '.join(node_tree)}\n{html}"
elif isinstance(msg, AIMessageChunk):
def truncate_middle(text, front=50, back=50):
if not text:
return ""
if len(text) <= front + back:
return text
return f"{text[:front]}{text[-back:]}".replace(
"\n", ""
) # remove new lines.
if not msg.content:
# logger.warning("*** No Message Chunk!")
yield f"### { " → ".join(node_tree)}\n{generate_processing_message()}\n<div class='intermediate-output'>{escape(truncate_middle(streamed_response))}</div>"
else:
# Stream intermediate messages with transparent style
# if node != final_node:
streamed_response += msg.content
yield f"### { ' → '.join(node_tree) }\n<div class='intermediate-output'>{escape(truncate_middle(streamed_response))}</div>"
# else:
# Buffer the final validated response instead of yielding
final_response += msg.content
if msg.tool_call_chunks:
for tool_call_chunk in msg.tool_call_chunks:
logger.debug("*** tool_call_chunk = ", tool_call_chunk)
if tool_call_chunk["id"] is not None:
tool_call_id = tool_call_chunk["id"]
if tool_call_id not in tool_calls_buffer:
tool_calls_buffer[tool_call_id] = {
"name": "",
"args_str": "",
"id": tool_call_id,
"type": "tool_call",
}
# Accumulate tool call name and arguments
if tool_call_chunk["name"] is not None:
tool_calls_buffer[tool_call_id]["name"] += tool_call_chunk[
"name"
]
if tool_call_chunk["args"] is not None:
tool_calls_buffer[tool_call_id][
"args_str"
] += tool_call_chunk["args"]
else:
logger.debug("message = ", type(msg), msg.content[:100])
full: str = escape(msg.content)
truncated = (
(full[:MAX_CONTENT] + "…") if len(full) > MAX_CONTENT else full
)
html = (
f"<div class='thinking-bubble'><em>🤔 {random.choice(thinking_verbs)} ...</em></div>"
f"<div style='opacity: 0.1'>"
f"<strong>Telling myself:</strong> {truncated or '...'}"
f"</div>"
)
yield f"### { " → ".join(node_tree)}\n{html}"
if getattr(msg, "tool_calls", []):
logger.info("ELSE::tool_calls = %s", msg.tool_calls)
node_tree[-1] = "✅"
end_time = time.time()
duration = end_time - start_time
final_response = (
f"\n{final_response}" f"\n\n⏱️ Processed in {duration:.2f} seconds"
)
buffer = f"### {' → '.join(node_tree)}\n"
yield buffer
for c in final_response:
buffer += c
yield buffer
await asyncio.sleep(0.0005)
logger.debug("************************************")
# Now, you can process the complete tool calls from the buffer
for tool_call_id, accumulated_tool_call in tool_calls_buffer.items():
# Attempt to parse arguments only if the 'args_str' isn't empty
if accumulated_tool_call["args_str"]:
try:
parsed_args = json.loads(accumulated_tool_call["args_str"])
logger.debug(f"Tool Name: {accumulated_tool_call['name']}")
logger.debug(f"Tool Arguments: {parsed_args}")
except json.JSONDecodeError:
logger.debug(
f"Partial arguments for tool {accumulated_tool_call['name']}: {accumulated_tool_call['args_str']}"
)
except asyncio.CancelledError:
logger.warning("⚠️ Request cancelled by user")
node_tree = end_node_tree(node_tree=node_tree)
yield (
f"### {' → '.join(node_tree)}"
"\n⚠️⚠️⚠️ Request cancelled by user"
"\nhere is what I got so far ...\n"
f"\n{streamed_response}"
)
# Important: re-raise if you want upstream to also know
# raise
return
except Exception as e:
logger.error("❌❌❌ Error processing request: %s", e)
traceback.print_exc()
node_tree = end_node_tree(node_tree=node_tree)
yield (
f"### { " → ".join(node_tree)}"
f"\n❌❌❌ Error processing request : {str(e)}"
"\nhere is what I got so far ...\n"
f"\n{streamed_response}"
)
return
def init_session():
return str(uuid.uuid4())
MAX_MESSAGES_IN_CONVERSATION = 5
async def limited_chat_wrapper(
message, history, thread_id, debug, preferred_language, count
):
# increment **after processing the message**
count += 1
# warn before reset
if count == MAX_MESSAGES_IN_CONVERSATION - 1:
yield [
{
"role": "system",
"content": "⚠️ You are allowed to ask one more follow-up. The next question will be considered a new conversation. Please wait ... processing your request ...",
}
], thread_id, count
await asyncio.sleep(1)
# reset
if count > MAX_MESSAGES_IN_CONVERSATION:
thread_id = init_session()
history = []
count = 1
yield [
{
"role": "system",
"content": "🔄 This is now considered a new question. Don't worry, your message shall still be processed! If I am giving irrelevant responses, you know why :-)",
}
], thread_id, count
await asyncio.sleep(1)
# normal flow: stream from your original chat_wrapper
final_chunk = []
async for chunk in chat_wrapper(
message, history, thread_id, debug, preferred_language
):
yield chunk, thread_id, count
final_chunk = chunk
# Simulating LLM Response
# for i in range(5):
# final_chunk += [{
# "role": "assistant",
# "content": f"Simulated LLM output {i+1}",
# }]
# yield final_chunk, thread_id, count
# await asyncio.sleep(0.25)
yield final_chunk, thread_id, count