Celes Renata
a72f336ad1
Multi-stage evidence-grounded inference architecture replacing the monolithic 9B model extraction pipeline. CPU-first specialist services handle routine extraction while the 9B vLLM model is preserved for semantic adjudication of ambiguous cases. Key components: - Capability-aware inference gateway (OpenAI-compatible + Ollama) - Endpoint registry with DB migrations and REST API - Sentence-aware document segmenter (property tests) - Deterministic financial parsing with offset integrity - Symbol resolution with ambiguity detection - Specialist service (GLiNER2, dynamic batching, K8s deployment) - Company-specific sentiment (FinBERT, calibration) - Retrieval-based novelty and duplicate detection - Confidence calibration pipeline - Deterministic routing engine (property tests) - 9B adjudication layer with VRAM gating - Stock-specific impact model (features, labels, baseline, trained) - Pipeline orchestrator (state machine, queues, leases, feature flags) - Bounded parallelism (async workers, semaphore, load shedding) - Observability (tracing, metrics, alerts) - Compatibility adapter (v3→v2 golden mapping tests) - Shadow/canary promotion framework - Active learning and fine-tuning pipeline Test results: 1,161 tests pass, ruff lint clean. All 282 spec tasks completed.
432 lines
15 KiB
Python
432 lines
15 KiB
Python
"""OpenAI-compatible inference client.
|
|
|
|
Uses httpx.AsyncClient directly (NOT the openai SDK). This keeps wire payloads
|
|
explicit, permits provider-specific extra_body, and simplifies redacted request
|
|
auditing.
|
|
|
|
Requirements: 2.1, 2.2, 2.3, 2.4, 2.5, 2.7, 2.9
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import logging
|
|
import os
|
|
import time
|
|
from typing import Any
|
|
|
|
import httpx
|
|
|
|
from services.shared.inference.models import (
|
|
ErrorCategory,
|
|
InferenceResult,
|
|
InferenceTarget,
|
|
StructuredGenerationRequest,
|
|
TokenUsage,
|
|
)
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
# Status codes that trigger retry
|
|
_RETRYABLE_STATUS_CODES = {429, 500, 502, 503, 504}
|
|
|
|
# Sensitive keys that must never appear in logs
|
|
_SENSITIVE_KEYS = frozenset({"authorization", "x-api-key", "api-key"})
|
|
|
|
|
|
def _resolve_auth_secret(secret_ref: str | None) -> str | None:
|
|
"""Resolve an authentication secret from environment variables.
|
|
|
|
The secret_ref is treated as an environment variable name.
|
|
Returns None if the ref is None or the env var is not set.
|
|
"""
|
|
if not secret_ref:
|
|
return None
|
|
return os.environ.get(secret_ref)
|
|
|
|
|
|
def _redact_headers(headers: dict[str, str]) -> dict[str, str]:
|
|
"""Return a copy of headers with sensitive values redacted."""
|
|
redacted = {}
|
|
for key, value in headers.items():
|
|
if key.lower() in _SENSITIVE_KEYS:
|
|
redacted[key] = "***REDACTED***"
|
|
else:
|
|
redacted[key] = value
|
|
return redacted
|
|
|
|
|
|
class OpenAICompatibleClient:
|
|
"""Client for OpenAI-compatible /v1/chat/completions endpoints.
|
|
|
|
Supports:
|
|
- Bearer and configurable authentication via runtime secret resolution
|
|
- Standard response_format.json_schema payloads
|
|
- Configurable vLLM structured_outputs extra-body payloads
|
|
- Explicit JSON-object and prompt-only fallback policies
|
|
- Retry on transient errors (5xx, 429, timeout)
|
|
- Local JSON schema revalidation
|
|
- Metadata capture (request_id, usage, finish_reason, retries)
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
target: InferenceTarget,
|
|
*,
|
|
http_client: httpx.AsyncClient | None = None,
|
|
) -> None:
|
|
self._target = target
|
|
self._owns_client = http_client is None
|
|
self._http = http_client or httpx.AsyncClient(timeout=target.timeout_seconds)
|
|
|
|
@property
|
|
def target(self) -> InferenceTarget:
|
|
"""The resolved inference target."""
|
|
return self._target
|
|
|
|
async def generate(
|
|
self,
|
|
request: StructuredGenerationRequest,
|
|
) -> InferenceResult:
|
|
"""Send a structured generation request and return the result.
|
|
|
|
Chooses the strongest structured-output mode the target supports:
|
|
1. json_schema — sends the actual schema with strict mode
|
|
2. json_object — only if target explicitly declares json_object capability
|
|
3. prompt_only — fallback when no structured output is available
|
|
"""
|
|
structured_mode = self._choose_structured_mode(request)
|
|
headers = self._build_headers()
|
|
body = self._build_body(request, structured_mode)
|
|
url = f"{self._target.base_url.rstrip('/')}/v1/chat/completions"
|
|
|
|
retries = 0
|
|
max_retries = self._target.max_retries
|
|
start_time = time.monotonic()
|
|
last_error: str | None = None
|
|
last_error_category: str | None = None
|
|
|
|
while True:
|
|
try:
|
|
response = await self._http.post(
|
|
url,
|
|
json=body,
|
|
headers=headers,
|
|
timeout=request.timeout_seconds,
|
|
)
|
|
except httpx.TimeoutException:
|
|
last_error = "Request timed out"
|
|
last_error_category = ErrorCategory.TIMEOUT
|
|
if retries < max_retries:
|
|
retries += 1
|
|
logger.warning(
|
|
"Timeout on attempt %d/%d to %s",
|
|
retries,
|
|
max_retries,
|
|
url,
|
|
)
|
|
continue
|
|
return self._error_result(
|
|
last_error,
|
|
last_error_category,
|
|
retries,
|
|
start_time,
|
|
structured_mode,
|
|
)
|
|
except httpx.ConnectError as exc:
|
|
last_error = f"Connection error: {exc}"
|
|
last_error_category = ErrorCategory.CONNECTION_ERROR
|
|
if retries < max_retries:
|
|
retries += 1
|
|
logger.warning(
|
|
"Connection error on attempt %d/%d to %s: %s",
|
|
retries,
|
|
max_retries,
|
|
url,
|
|
exc,
|
|
)
|
|
continue
|
|
return self._error_result(
|
|
last_error,
|
|
last_error_category,
|
|
retries,
|
|
start_time,
|
|
structured_mode,
|
|
)
|
|
|
|
# Handle retryable HTTP status codes
|
|
if response.status_code in _RETRYABLE_STATUS_CODES:
|
|
if response.status_code == 429:
|
|
last_error_category = ErrorCategory.RATE_LIMIT
|
|
last_error = "Rate limited (429)"
|
|
else:
|
|
last_error_category = ErrorCategory.SERVER_ERROR
|
|
last_error = f"Server error ({response.status_code})"
|
|
|
|
if retries < max_retries:
|
|
retries += 1
|
|
logger.warning(
|
|
"HTTP %d on attempt %d/%d to %s",
|
|
response.status_code,
|
|
retries,
|
|
max_retries,
|
|
url,
|
|
)
|
|
continue
|
|
return self._error_result(
|
|
last_error,
|
|
last_error_category,
|
|
retries,
|
|
start_time,
|
|
structured_mode,
|
|
)
|
|
|
|
# Handle non-retryable errors
|
|
if response.status_code == 401 or response.status_code == 403:
|
|
return self._error_result(
|
|
f"Authentication failed ({response.status_code})",
|
|
ErrorCategory.AUTHENTICATION,
|
|
retries,
|
|
start_time,
|
|
structured_mode,
|
|
)
|
|
|
|
if response.status_code >= 400:
|
|
return self._error_result(
|
|
f"Client error ({response.status_code})",
|
|
ErrorCategory.INVALID_RESPONSE,
|
|
retries,
|
|
start_time,
|
|
structured_mode,
|
|
)
|
|
|
|
# Success path
|
|
break
|
|
|
|
latency_ms = int((time.monotonic() - start_time) * 1000)
|
|
return self._parse_response(
|
|
response,
|
|
request,
|
|
structured_mode,
|
|
retries,
|
|
latency_ms,
|
|
)
|
|
|
|
def _choose_structured_mode(
|
|
self, request: StructuredGenerationRequest
|
|
) -> str:
|
|
"""Choose the strongest structured-output mode available."""
|
|
caps = self._target.capabilities
|
|
|
|
if request.json_schema and caps.json_schema:
|
|
return "json_schema"
|
|
if request.json_schema and caps.json_object:
|
|
return "json_object"
|
|
if request.json_schema:
|
|
# Schema requested but endpoint supports neither — prompt-only fallback
|
|
return "prompt_only"
|
|
return "none"
|
|
|
|
def _build_headers(self) -> dict[str, str]:
|
|
"""Build request headers including authentication."""
|
|
headers: dict[str, str] = {
|
|
"Content-Type": "application/json",
|
|
}
|
|
|
|
# Add extra headers from target config
|
|
headers.update(self._target.extra_headers)
|
|
|
|
# Resolve auth secret
|
|
secret = _resolve_auth_secret(self._target.auth_secret_ref)
|
|
if secret:
|
|
scheme = self._target.auth_scheme.lower()
|
|
if scheme == "bearer":
|
|
headers["Authorization"] = f"Bearer {secret}"
|
|
else:
|
|
# Custom auth scheme (e.g., "X-API-Key: <value>")
|
|
headers[self._target.auth_scheme] = secret
|
|
|
|
return headers
|
|
|
|
def _build_body(
|
|
self,
|
|
request: StructuredGenerationRequest,
|
|
structured_mode: str,
|
|
) -> dict[str, Any]:
|
|
"""Build the request body for /v1/chat/completions."""
|
|
messages = [
|
|
{"role": msg.role, "content": msg.content}
|
|
for msg in request.messages
|
|
]
|
|
|
|
body: dict[str, Any] = {
|
|
"model": self._target.model,
|
|
"messages": messages,
|
|
"temperature": request.temperature,
|
|
"max_tokens": request.max_output_tokens,
|
|
}
|
|
|
|
# Add seed if the target supports it
|
|
if self._target.capabilities.seed and request.seed is not None:
|
|
body["seed"] = request.seed
|
|
|
|
# Add response_format based on chosen mode
|
|
if structured_mode == "json_schema" and request.json_schema:
|
|
body["response_format"] = {
|
|
"type": "json_schema",
|
|
"json_schema": {
|
|
"name": request.json_schema.get("title", "response"),
|
|
"strict": True,
|
|
"schema": request.json_schema,
|
|
},
|
|
}
|
|
elif structured_mode == "json_object":
|
|
body["response_format"] = {"type": "json_object"}
|
|
|
|
# Add extra_body from target config (vLLM structured_outputs, etc.)
|
|
if self._target.extra_body:
|
|
body.update(self._target.extra_body)
|
|
|
|
return body
|
|
|
|
def _parse_response(
|
|
self,
|
|
response: httpx.Response,
|
|
request: StructuredGenerationRequest,
|
|
structured_mode: str,
|
|
retries: int,
|
|
latency_ms: int,
|
|
) -> InferenceResult:
|
|
"""Parse a successful response into InferenceResult."""
|
|
# Extract request ID from response headers
|
|
request_id = response.headers.get("x-request-id")
|
|
|
|
try:
|
|
data = response.json()
|
|
except (json.JSONDecodeError, ValueError):
|
|
return InferenceResult(
|
|
content="",
|
|
endpoint_id=self._target.endpoint_id,
|
|
deployment_id=self._target.deployment_id,
|
|
model=self._target.model,
|
|
protocol=self._target.protocol,
|
|
structured_mode=structured_mode, # type: ignore[arg-type]
|
|
latency_ms=latency_ms,
|
|
retries=retries,
|
|
request_id=request_id,
|
|
error="Invalid JSON in response body",
|
|
error_category=ErrorCategory.INVALID_RESPONSE,
|
|
)
|
|
|
|
# Extract content from choices
|
|
choices = data.get("choices", [])
|
|
if not choices:
|
|
return InferenceResult(
|
|
content="",
|
|
endpoint_id=self._target.endpoint_id,
|
|
deployment_id=self._target.deployment_id,
|
|
model=self._target.model,
|
|
protocol=self._target.protocol,
|
|
structured_mode=structured_mode, # type: ignore[arg-type]
|
|
latency_ms=latency_ms,
|
|
retries=retries,
|
|
request_id=request_id,
|
|
error="Empty choices in response",
|
|
error_category=ErrorCategory.INVALID_RESPONSE,
|
|
)
|
|
|
|
message = choices[0].get("message", {})
|
|
content = message.get("content", "")
|
|
finish_reason = choices[0].get("finish_reason")
|
|
|
|
# Extract usage metadata
|
|
usage_data = data.get("usage", {})
|
|
input_tokens = usage_data.get("prompt_tokens")
|
|
output_tokens = usage_data.get("completion_tokens")
|
|
total_tokens = usage_data.get("total_tokens")
|
|
|
|
# Parse JSON content if structured output was requested
|
|
parsed: dict[str, Any] | None = None
|
|
schema_valid: bool | None = None
|
|
|
|
if structured_mode in ("json_schema", "json_object") and content:
|
|
try:
|
|
parsed = json.loads(content)
|
|
except (json.JSONDecodeError, ValueError):
|
|
# Content is not valid JSON
|
|
schema_valid = False
|
|
|
|
# Validate parsed JSON against supplied schema
|
|
if parsed is not None and request.json_schema:
|
|
schema_valid = self._validate_schema(parsed, request.json_schema)
|
|
|
|
return InferenceResult(
|
|
content=content,
|
|
parsed=parsed,
|
|
endpoint_id=self._target.endpoint_id,
|
|
deployment_id=self._target.deployment_id,
|
|
model=self._target.model,
|
|
protocol=self._target.protocol,
|
|
structured_mode=structured_mode, # type: ignore[arg-type]
|
|
latency_ms=latency_ms,
|
|
usage=TokenUsage(
|
|
input_tokens=input_tokens,
|
|
output_tokens=output_tokens,
|
|
total_tokens=total_tokens,
|
|
),
|
|
request_id=request_id,
|
|
finish_reason=finish_reason,
|
|
retries=retries,
|
|
schema_valid=schema_valid,
|
|
)
|
|
|
|
def _validate_schema(
|
|
self, data: dict[str, Any], schema: dict[str, Any]
|
|
) -> bool:
|
|
"""Revalidate parsed JSON locally against the supplied schema.
|
|
|
|
Returns True if valid, False otherwise.
|
|
"""
|
|
try:
|
|
import jsonschema
|
|
|
|
jsonschema.validate(instance=data, schema=schema)
|
|
return True
|
|
except Exception:
|
|
logger.debug("Schema validation failed for response data")
|
|
return False
|
|
|
|
def _error_result(
|
|
self,
|
|
error: str,
|
|
error_category: str,
|
|
retries: int,
|
|
start_time: float,
|
|
structured_mode: str,
|
|
) -> InferenceResult:
|
|
"""Build an error InferenceResult."""
|
|
latency_ms = int((time.monotonic() - start_time) * 1000)
|
|
return InferenceResult(
|
|
content="",
|
|
endpoint_id=self._target.endpoint_id,
|
|
deployment_id=self._target.deployment_id,
|
|
model=self._target.model,
|
|
protocol=self._target.protocol,
|
|
structured_mode=structured_mode, # type: ignore[arg-type]
|
|
latency_ms=latency_ms,
|
|
retries=retries,
|
|
error=error,
|
|
error_category=error_category,
|
|
)
|
|
|
|
async def close(self) -> None:
|
|
"""Close the underlying HTTP client if we own it."""
|
|
if self._owns_client:
|
|
await self._http.aclose()
|
|
|
|
def __repr__(self) -> str:
|
|
return (
|
|
f"OpenAICompatibleClient("
|
|
f"model={self._target.model!r}, "
|
|
f"base_url={self._target.base_url!r})"
|
|
)
|