Hume EVI + LangGraph Integration#

Architecture#

Single LangGraph StateGraph with interrupt/resume:

text

receive_call → verify_pin → select_persona → create_hume_config → generate_twiml
    → await_call_end [INTERRUPT] → fetch_transcript → analyze → coach → store → END

The interrupt boundary separates pre-call (synchronous) from post-call (webhook-triggered).

Critical Patterns#

1. Interrupt/Resume for Async Calls#

python

from langgraph.types import interrupt, Command

def await_call_end(state):
    resume_data = interrupt({"reason": "waiting_for_webhook"})
    return {**state, "chat_id": resume_data["chat_id"]}

# In webhook handler:
graph.invoke(Command(resume={"chat_id": "xxx"}), config)

2. Hume Config Creation#

Create dynamic EVI configs per call. Set temperature low (0.6) to prevent default enthusiasm:

python

request_body = {
    "evi_version": "3",
    "name": f"Session-{persona_name}-{timestamp}",
    "prompt": {"text": voice_prompt},
    "voice": {"provider": "HUME_AI", "name": "KORA"},  # or "ITO" for male
    "language_model": {
        "model_provider": "OPEN_AI",
        "model_resource": "gpt-4o-mini",
        "temperature": 0.6,  # CRITICAL: default is too warm/eager
    },
    "event_messages": {"on_new_chat": {"enabled": True, "text": first_message}},
    "webhooks": [{"events": ["chat_ended"], "url": webhook_url}],
}
resp = httpx.post("https://api.hume.ai/v0/evi/configs", json=request_body, headers=headers)

3. TwiML Redirect (not Stream)#

python

twiml = f'''<?xml version="1.0" encoding="UTF-8"?>
<Response>
    <Say voice="Polly.Matthew">Connecting now.</Say>
    <Redirect>https://api.hume.ai/v0/evi/twilio?config_id={config_id}&amp;api_key={api_key}</Redirect>
</Response>'''

Use & not & — this is inside XML.

4. Transcript Fetching (⚠️ Known Bug Zone)#

Hume's /chats/{id}/events returns 404. Must use chat_groups:

python

# Step 1: Get chat_group_id
chat_resp = httpx.get(f"https://api.hume.ai/v0/evi/chats/{chat_id}", headers=headers)
chat_group_id = chat_resp.json().get("chat_group_id")

# Step 2: Fetch events via chat_group
events_resp = httpx.get(
    f"https://api.hume.ai/v0/evi/chat_groups/{chat_group_id}/events",
    headers=headers, params={"page_size": 100}
)
events = events_resp.json().get("events_page", [])

Field names are snake_case: message_text, emotion_features (not messageText).

5. Emotion Extraction#

python

for msg in messages:
    ef = msg.get("emotion_features")  # dict of ~48 emotions with float scores
    if ef and msg.get("role") == "USER":  # USER = the human caller
        top = sorted(ef.items(), key=lambda x: x[1], reverse=True)[:5]
        emotion_timeline.append({"turn": n, "text": text, "top_emotions": dict(top)})

6. Webhook Session Resolution#

Hume chat_ended webhook does NOT include call_sid. Use config_id mapping:

python

config_to_thread: dict[str, str] = {}  # hume_config_id → langgraph_thread_id

# On config creation:
config_to_thread[config_id] = thread_id

# On webhook:
thread_id = config_to_thread.pop(body["config_id"])

Prevention Rules#

See references/bug-prevention.md for the full bug registry and prevention checklist.