Files
eventflow/backend/fastapi_app/services/llm_extract.py

340 lines
11 KiB
Python

import json
import os
import subprocess
import time
import urllib.error
import urllib.request
from typing import Any
from openai import OpenAI
def _client() -> OpenAI:
base_url = os.environ.get("LLM_BASE_URL")
api_key = os.environ.get("LLM_API_KEY")
if not api_key:
raise RuntimeError("LLM_API_KEY is not set")
# Keep LLM calls bounded; some gateways hang on upstream issues.
timeout_s = float(os.environ.get("LLM_TIMEOUT_S", "30"))
# OpenAI SDK uses base_url for OpenAI-compatible gateways.
return OpenAI(api_key=api_key, base_url=base_url, timeout=timeout_s)
SYSTEM_PROMPT = (
"You are a finance event extraction engine. Extract a single structured event. "
"Be conservative: if uncertain, set impact_direction=0 and confidence low. "
"Output MUST be valid JSON only (no markdown, no commentary)."
)
def _coerce_result(obj: dict[str, Any]) -> dict[str, Any]:
# Keep the contract stable for downstream DB writes.
out: dict[str, Any] = {}
out["event_type"] = str(obj.get("event_type") or "unknown")
region = obj.get("region")
if region is not None and str(region).strip() != "":
out["region"] = str(region)
entities = obj.get("entities")
if isinstance(entities, list):
out["entities"] = [str(x) for x in entities if str(x).strip()]
else:
out["entities"] = []
out["summary_zh"] = str(obj.get("summary_zh") or "")
out["summary_en"] = str(obj.get("summary_en") or "")
# -1/0/1 only
try:
impact_direction = int(obj.get("impact_direction"))
except Exception:
impact_direction = 0
if impact_direction not in (-1, 0, 1):
impact_direction = 0
out["impact_direction"] = impact_direction
# 0..1
try:
confidence = float(obj.get("confidence"))
except Exception:
confidence = 0.0
if confidence < 0.0:
confidence = 0.0
if confidence > 1.0:
confidence = 1.0
out["confidence"] = confidence
evidence = obj.get("evidence")
if isinstance(evidence, list):
out["evidence"] = [str(x) for x in evidence if str(x).strip()]
else:
out["evidence"] = []
return out
def _responses_output_text(resp_obj: dict[str, Any]) -> str:
txt = resp_obj.get("output_text")
if isinstance(txt, str) and txt.strip():
return txt
# Fallback for gateways that only return structured `output` blocks.
out = resp_obj.get("output")
if isinstance(out, list):
chunks: list[str] = []
for item in out:
if not isinstance(item, dict):
continue
content = item.get("content")
if not isinstance(content, list):
continue
for block in content:
if not isinstance(block, dict):
continue
if block.get("type") in ("output_text", "text"):
val = block.get("text")
if isinstance(val, str) and val.strip():
chunks.append(val)
if chunks:
return "\n".join(chunks)
return "{}"
def _find_json_object(text: str) -> str | None:
start = text.find("{")
while start != -1:
depth = 0
in_str = False
esc = False
for i in range(start, len(text)):
ch = text[i]
if in_str:
if esc:
esc = False
elif ch == "\\":
esc = True
elif ch == '"':
in_str = False
continue
if ch == '"':
in_str = True
elif ch == "{":
depth += 1
elif ch == "}":
depth -= 1
if depth == 0:
candidate = text[start : i + 1]
try:
json.loads(candidate)
return candidate
except Exception:
break
start = text.find("{", start + 1)
return None
def _responses_create_http(model: str, user_payload: dict[str, Any]) -> dict[str, Any]:
base_url = (os.environ.get("LLM_BASE_URL") or "").rstrip("/")
api_key = os.environ.get("LLM_API_KEY")
if not base_url:
raise RuntimeError("LLM_BASE_URL is not set")
if not api_key:
raise RuntimeError("LLM_API_KEY is not set")
timeout_s = float(os.environ.get("LLM_TIMEOUT_S", "30"))
body = {
"model": model,
"stream": False,
"max_output_tokens": int(os.environ.get("LLM_MAX_OUTPUT_TOKENS", "512")),
"input": [
{
"role": "user",
"content": [
{
"type": "text",
"text": (
SYSTEM_PROMPT
+ "\n\nReturn valid JSON object only.\n\nInput:\n"
+ json.dumps(user_payload, ensure_ascii=True)
),
}
],
}
],
}
req = urllib.request.Request(
url=f"{base_url}/responses",
data=json.dumps(body).encode("utf-8"),
headers={
"Content-Type": "application/json",
"Authorization": f"Bearer {api_key}",
},
method="POST",
)
try:
with urllib.request.urlopen(req, timeout=timeout_s) as resp:
raw = resp.read().decode("utf-8", errors="replace")
except urllib.error.HTTPError as e:
detail = e.read().decode("utf-8", errors="replace")
raise RuntimeError(f"http_{e.code}: {detail[:5000]}") from e
try:
return json.loads(raw)
except json.JSONDecodeError as e:
raise RuntimeError(f"invalid_gateway_json: {raw[:1000]}") from e
def _codex_exec_fallback(model: str, user_payload: dict[str, Any]) -> dict[str, Any]:
codex_bin = os.environ.get("CODEX_BIN", os.path.expanduser("~/.local/npm-global/bin/codex"))
codex_timeout_s = float(os.environ.get("CODEX_FALLBACK_TIMEOUT_S", "90"))
codex_cwd = os.environ.get("CODEX_FALLBACK_CWD", "/home/openclaw")
prompt = (
"Return ONLY one valid JSON object with keys: "
"event_type,region,entities,summary_zh,summary_en,impact_direction,confidence,evidence.\n"
"No markdown, no extra text.\n"
f"Input:\n{json.dumps(user_payload, ensure_ascii=True)}"
)
proc = subprocess.run(
[
codex_bin,
"exec",
"-m",
model,
"-s",
"read-only",
"-C",
codex_cwd,
"--skip-git-repo-check",
prompt,
],
capture_output=True,
text=True,
timeout=codex_timeout_s,
check=False,
)
combined = (proc.stdout or "") + "\n" + (proc.stderr or "")
candidate = _find_json_object(combined)
if not candidate:
raise RuntimeError(f"codex_no_json: {combined[:1000]}")
return {"output_text": candidate}
def extract_event(title: str, content: str, lang_hint: str | None = None) -> dict[str, Any]:
"""Return a strict JSON object matching our event schema.
Supports both OpenAI-compatible APIs:
- chat.completions (response_format=json_object)
- responses (text.format=json_object)
"""
model = os.environ.get("LLM_MODEL", "gpt-5.2")
wire_api = (os.environ.get("LLM_WIRE_API") or os.environ.get("WIRE_API") or "chat").strip().lower()
user = {
"title": title,
"content": content,
"lang_hint": lang_hint,
"required_output": {
"event_type": "string (macro/policy/war/sanctions/earnings/commodity/supply_shock/rate/inflation/...) ",
"region": "string optional (CN/US/EU/JP/Global/...)",
"entities": "array of strings (companies/countries/commodities/indices)",
"summary_zh": "string",
"summary_en": "string",
"impact_direction": "-1|0|1 (market risk sentiment for affected assets)",
"confidence": "float 0..1",
"evidence": "array of strings (short quotes or key facts; URLs if present)",
},
}
c = _client() if wire_api not in ("responses", "response") else None
max_attempts = int(os.environ.get("LLM_MAX_ATTEMPTS", "3"))
backoff_s = float(os.environ.get("LLM_RETRY_BACKOFF_S", "1.0"))
last_exc: Exception | None = None
resp = None
for attempt in range(1, max_attempts + 1):
try:
if wire_api in ("responses", "response"):
# CCH is more reliable with direct /v1/responses HTTP + stream=false.
resp = _responses_create_http(model=model, user_payload=user)
else:
resp = c.chat.completions.create(
model=model,
messages=[
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": json.dumps(user, ensure_ascii=True)},
],
temperature=0.2,
response_format={"type": "json_object"},
)
last_exc = None
break
except Exception as e:
# Gateways can intermittently throw 5xx/connection errors. Retry a few times.
last_exc = e
# Optional escape hatch: fallback to local Codex CLI when CCH responses are unstable.
use_codex_fallback = os.environ.get("LLM_FALLBACK_CODEX", "1").lower() in ("1", "true", "yes")
if use_codex_fallback and wire_api in ("responses", "response"):
try:
resp = _codex_exec_fallback(model=model, user_payload=user)
last_exc = None
break
except Exception as fallback_exc:
last_exc = fallback_exc
if attempt < max_attempts:
time.sleep(backoff_s * (2 ** (attempt - 1)))
continue
break
if resp is None:
return {
"model": model,
"ok": False,
"error": "llm_failed",
"exc": type(last_exc).__name__ if last_exc else "Unknown",
"msg": str(last_exc)[:5000] if last_exc else "",
}
if wire_api in ("responses", "response"):
txt = _responses_output_text(resp if isinstance(resp, dict) else {})
else:
txt = resp.choices[0].message.content or "{}"
try:
obj = json.loads(txt)
except json.JSONDecodeError:
# Return raw text for debugging, but still keep a predictable envelope.
return {
"model": model,
"ok": False,
"error": "invalid_json",
"raw": txt,
}
if not isinstance(obj, dict):
return {
"model": model,
"ok": False,
"error": "non_object_json",
"raw": txt,
}
return {
"model": model,
"ok": True,
"event": _coerce_result(obj),
}