avrik-server 1.0.0__py3-none-any.whl

avrik_server/__init__.py

@@ -0,0 +1,4 @@
+"""Avrik server: readout, calibration, /v1/systemone API and inference backends.
+
+Skeleton only; see server/ARCHITECTURE.md.
+"""

avrik_server/api/__init__.py

@@ -0,0 +1 @@
+"""The /v1/systemone HTTP service (Step 2.5)."""

avrik_server/api/__main__.py

@@ -0,0 +1,77 @@
+"""``python -m avrik_server.api --config <server.yaml>``: serve ``/v1/systemone``.
+
+Loads the config, the pinned tokenizer (local files only, no network), the backend and
+the calibration file; refuses to start on a mismatched calibration; binds to the
+configured host (loopback unless API keys are set, research R11).
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from collections.abc import Sequence
+from pathlib import Path
+from typing import Any
+
+from avrik_server.config import ServerConfig, load_config
+
+
+def build_backend(config: ServerConfig) -> Any:  # pragma: no cover - loads real backends
+    """The configured backend, async."""
+    b = config.backend
+    if b.kind == "vllm-gpu":
+        from avrik_server.backends.vllm_http import AsyncVllmHttpBackend
+
+        return AsyncVllmHttpBackend(
+            str(b.url),
+            model=b.served_model,
+            precision=b.precision,
+            timeout=b.timeout,
+            full_vocab=b.full_vocab,
+            sequential_shards=b.sequential_shards,
+            shard_size=b.shard_size,
+        )
+    from avrik_server.backends.threaded import ThreadedBackend
+
+    if b.kind == "llamacpp-cpu":
+        from avrik_server.backends.llamacpp import LlamaCppBackend
+
+        return ThreadedBackend(LlamaCppBackend(str(b.gguf), n_ctx=b.max_model_len))
+    from avrik_server.backends.transformers_ref import TransformersBackend
+
+    dtype = "float32" if b.precision == "fp32" else "bfloat16"
+    return ThreadedBackend(TransformersBackend(config.model_dir, dtype=dtype))
+
+
+def main(argv: Sequence[str] | None = None) -> int:  # pragma: no cover - starts a server
+    """Entry point."""
+    parser = argparse.ArgumentParser(prog="python -m avrik_server.api")
+    parser.add_argument("--config", type=Path, required=True)
+    args = parser.parse_args(argv)
+    import uvicorn
+
+    from avrik_server.api.app import create_app
+    from avrik_server.calibration.file import CalibrationError, load
+    from avrik_server.engine.decide import Engine
+    from avrik_server.readout.smoke import _hf_tokenizer
+
+    config = load_config(args.config)
+    try:
+        engine = Engine(
+            config, _hf_tokenizer(config.model_dir), build_backend(config), load(config.calibration)
+        )
+    except CalibrationError as error:
+        print(f"refused: {error}", file=sys.stderr)
+        return 3
+    uvicorn.run(
+        create_app(engine),
+        host=config.serving.host,
+        port=config.serving.port,
+        log_level="info",
+        access_log=False,
+    )
+    return 0
+
+
+if __name__ == "__main__":  # pragma: no cover
+    sys.exit(main())

avrik_server/api/app.py

@@ -0,0 +1,143 @@
+"""The HTTP service (build guide Step 2.5; contracts/systemone-api.md; research R11).
+
+``create_app(engine)`` builds the ASGI app around an ``Engine``:
+
+* ``POST /v1/systemone``: auth (when keys are set), per-key buckets, the bounded decide
+  queue, then the raw body is size-checked (413), parsed (400) and handed to the engine,
+  whose refusals become stable JSON errors;
+* ``GET /health``, ``/v1/models``, ``/openapi.json``, ``/llms.txt``: public discovery;
+* every response carries the RateLimit headers; unknown paths answer a JSON 404.
+"""
+
+from __future__ import annotations
+
+import json
+import re
+import uuid
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from fastapi import FastAPI, Request, Response
+from fastapi.responses import JSONResponse, PlainTextResponse
+from starlette.exceptions import HTTPException as StarletteHTTPException
+
+from avrik_server.api import discovery
+from avrik_server.api.errors import ApiError
+from avrik_server.api.ratelimit import Limiter, RetryAfter
+from avrik_server.engine.decide import Engine
+
+REQUEST_ID = re.compile(r"[A-Za-z0-9._-]{1,64}")
+
+
+async def read_capped(request: Request, limit: int) -> bytes:
+    """The body, refused as soon as it passes ``limit`` bytes (chunked bodies too; T056 B1).
+
+    Raises:
+        ApiError: ``body_too_large``.
+    """
+    chunks: list[bytes] = []
+    size = 0
+    async for chunk in request.stream():
+        size += len(chunk)
+        if size > limit:
+            raise ApiError("body_too_large", f"bodies are limited to {limit} bytes")
+        chunks.append(chunk)
+    return b"".join(chunks)
+
+
+def create_app(engine: Engine, limiter: Limiter | None = None) -> FastAPI:
+    """The ASGI app serving ``engine``."""
+    app = FastAPI(
+        title="Avrik decision API",
+        version="1.0.0",
+        openapi_url=None,
+        docs_url=None,
+        redoc_url=None,
+    )
+    app.state.engine = engine
+    limits = limiter or Limiter(engine.config.serving)
+    max_body = engine.config.limits.max_body_bytes
+
+    def error_response(error: ApiError, who: str | None = None) -> JSONResponse:
+        response = JSONResponse(error.body(), status_code=error.status)
+        if isinstance(error, RetryAfter):
+            response.headers["Retry-After"] = str(error.retry_after)
+        response.headers.update(limits.headers(who))
+        return response
+
+    @app.exception_handler(ApiError)
+    async def api_error(_: Request, error: ApiError) -> JSONResponse:
+        return error_response(error)
+
+    @app.exception_handler(StarletteHTTPException)
+    async def http_error(_: Request, error: StarletteHTTPException) -> JSONResponse:
+        code = "not_found" if error.status_code == 404 else "invalid_request"
+        body = {"error": {"code": code, "message": str(error.detail), "field": None}}
+        response = JSONResponse(body, status_code=error.status_code)
+        response.headers.update(limits.headers(None))
+        return response
+
+    @app.middleware("http")
+    async def rate_headers(
+        request: Request, call_next: Callable[[Request], Awaitable[Response]]
+    ) -> Response:
+        response = await call_next(request)
+        for name, value in limits.headers(getattr(request.state, "who", None)).items():
+            response.headers.setdefault(name, value)
+        given = request.headers.get("x-request-id", "")
+        safe = bool(REQUEST_ID.fullmatch(given))  # echo a client id only if it is plain
+        response.headers.setdefault("X-Request-Id", given if safe else uuid.uuid4().hex)
+        return response
+
+    @app.post("/v1/systemone")
+    async def systemone(request: Request) -> JSONResponse:
+        client = request.client.host if request.client else "unknown"
+        who = limits.identify(request.headers.get("authorization"), client)
+        request.state.who = who
+        try:
+            limits.charge(who)
+        except ApiError as error:
+            return error_response(error, who)
+        declared = request.headers.get("content-length")
+        if declared is not None and declared.isdigit() and int(declared) > max_body:
+            raise ApiError("body_too_large", f"bodies are limited to {max_body} bytes")
+        raw = await read_capped(request, max_body)
+        try:
+            body: Any = json.loads(raw)
+        except ValueError as error:
+            reason = getattr(error, "msg", "undecodable bytes")
+            raise ApiError("bad_json", f"the body is not valid JSON: {reason}") from error
+        try:
+            await limits.acquire()
+        except ApiError as error:
+            return error_response(error, who)
+        try:
+            result = await engine.decide(body, tenant=who)
+        finally:
+            limits.release()
+        response = JSONResponse(result)
+        response.headers.update(limits.headers(who))
+        return response
+
+    @app.get("/health")
+    async def health() -> JSONResponse:
+        ok = await engine.backend.health()
+        return JSONResponse(discovery.health(engine, ok), status_code=200 if ok else 503)
+
+    @app.get("/v1/models")
+    async def models() -> JSONResponse:
+        return JSONResponse(discovery.models(engine))
+
+    spec: dict[str, Any] = {}
+
+    @app.get("/openapi.json")
+    async def openapi() -> JSONResponse:
+        if not spec:  # built once (T056 N1)
+            spec.update(discovery.openapi(app))
+        return JSONResponse(spec)
+
+    @app.get("/llms.txt")
+    async def llms() -> PlainTextResponse:
+        return PlainTextResponse(discovery.LLMS_TXT)
+
+    return app

avrik_server/api/discovery.py

@@ -0,0 +1,129 @@
+"""Discovery: ``/health``, ``/v1/models``, ``/openapi.json`` and ``/llms.txt`` (FR-019).
+
+``/v1/models`` keeps Jev's ``{"models": [{name, description, release_date}]}`` shape with
+an additive ``avrik`` object; ``/health`` and ``/v1/models`` report the same hashes the
+audit block carries. The OpenAPI document (3.1) gets the request schema injected, since
+the route reads the raw body to keep stable error codes.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from fastapi import FastAPI
+
+from avrik_server.api.schemas import ErrorBody, SystemOneRequest, SystemOneResponse
+from avrik_server.engine.decide import Engine
+
+LLMS_TXT = """# Avrik decision API
+
+> One forward pass, no generated text: state + typed questions in, calibrated
+> probability distributions out. Shape-compatible with TypeSafe's /v1/systemone.
+
+## Endpoint
+
+POST /v1/systemone with {"model", "state", "questions"}.
+- state: text, a JSON object, or an array; a batch is an array of {id, text, questions}.
+- questions: {name: {type, instructions, criteria}}; types noul (yes/no), choice
+  (2 to max_options options), score (2-10 levels); Avrik extension multi_label
+  ({labels: [...]}, one yes/no per label, default threshold 0.7).
+- Optional Avrik fields: abstain_option per question (off by default), options
+  {near_tie_margin, near_tie_reask, correction}, images (base64, when enabled).
+
+## Answers
+
+Jev fields per type (choice + probabilities + confidence; noul; score + legend), plus
+top, margin, abstain, avrik_confidence, unknown_probability (abstain offered) and an
+eight-field audit block (model, adapter, calibration, template, allowed set, input
+tokens, backend, precision).
+
+Confidence measures which offered option fits best, not whether any fits: offer a
+none-of-the-above option when that can be the answer.
+
+## Discovery
+
+GET /v1/models (public), GET /health, GET /openapi.json.
+"""
+
+
+def models(engine: Engine) -> dict[str, Any]:
+    """The ``/v1/models`` body."""
+    config = engine.config
+    return {
+        "models": [
+            {
+                "name": config.model_name,
+                "description": config.description or f"Frozen {config.pin['repo']}",
+                "release_date": config.release_date or None,
+                "avrik": {
+                    "model_hash": engine.model_hash,
+                    "adapter_hash": engine.adapter_hash,
+                    "calibration_version": engine.calibration.version,
+                    "prompt_template_sha256": engine.template,
+                    "backend": engine.backend.name,
+                    "precision": engine.backend.precision,
+                    "max_options": engine.max_options,
+                    "max_questions": config.limits.max_questions,
+                    "max_context_tokens": config.backend.max_model_len - 1,
+                    "images": config.images,
+                },
+            }
+        ]
+    }
+
+
+def health(engine: Engine, ok: bool) -> dict[str, Any]:
+    """The ``/health`` body."""
+    if not ok:
+        return {"status": "backend_unavailable"}
+    return {
+        "status": "ok",
+        "model_hash": engine.model_hash,
+        "calibration_version": engine.calibration.version,
+        "prompt_template_sha256": engine.template,
+        "backend": engine.backend.name,
+        "precision": engine.backend.precision,
+    }
+
+
+def openapi(app: FastAPI) -> dict[str, Any]:
+    """OpenAPI 3.1 with the ``/v1/systemone`` request and response schemas."""
+    from fastapi.openapi.utils import get_openapi
+
+    spec = get_openapi(title=app.title, version=app.version, routes=app.routes)
+    components = spec.setdefault("components", {}).setdefault("schemas", {})
+    for model in (SystemOneRequest, SystemOneResponse, ErrorBody):
+        schema = model.model_json_schema(ref_template="#/components/schemas/{model}")
+        components.update(schema.pop("$defs", {}))
+        components[model.__name__] = schema
+    post = spec["paths"]["/v1/systemone"]["post"]
+    post["requestBody"] = {
+        "required": True,
+        "content": {
+            "application/json": {"schema": {"$ref": "#/components/schemas/SystemOneRequest"}}
+        },
+    }
+    error = {
+        "content": {"application/json": {"schema": {"$ref": "#/components/schemas/ErrorBody"}}}
+    }
+    post["responses"] = {
+        "200": {
+            "description": "Answers",
+            "content": {
+                "application/json": {"schema": {"$ref": "#/components/schemas/SystemOneResponse"}}
+            },
+        },
+        **{
+            code: {"description": text, **error}
+            for code, text in (
+                ("400", "Not JSON"),
+                ("401", "Bad key"),
+                ("413", "Body too large"),
+                ("422", "Refused, with a stable code"),
+                ("429", "Rate limited or at capacity"),
+                ("502", "Backend error"),
+                ("503", "Backend unavailable"),
+            )
+        },
+    }
+    return spec

avrik_server/api/errors.py

@@ -0,0 +1,57 @@
+"""Stable error codes and the error body (contracts/systemone-api.md "Errors").
+
+Every refusal is ``{"error": {"code", "message", "field"}}`` with a code from the
+systemone-contract skill (plus ``invalid_request``, ``readout_mass``, ``backend_error``,
+``body_too_large``, ``bad_json`` and ``not_found`` for cases the skill's list does not
+name). Capacity messages carry the Decision Index's markers ("options per choice",
+"maximum context length") so its http engine counts them as unsupported, not failed.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+STATUS: dict[str, int] = {
+    "bad_json": 400,
+    "unauthorized": 401,
+    "not_found": 404,
+    "body_too_large": 413,
+    "invalid_request": 422,
+    "too_few_options": 422,
+    "too_many_options": 422,
+    "context_too_large": 422,
+    "too_many_questions": 422,
+    "bad_question_type": 422,
+    "images_unsupported": 422,
+    "image_too_large": 422,
+    "model_unavailable": 422,
+    "readout_mass": 422,
+    "rate_limit_minute": 429,
+    "rate_limit_day": 429,
+    "request_spending_limit": 429,
+    "backend_error": 502,
+    "backend_unavailable": 503,
+}
+
+
+class ApiError(Exception):
+    """A refusal with a stable code.
+
+    Args:
+        code: A key of ``STATUS``.
+        message: Human-readable reason (never request content beyond field names).
+        field: JSON path of the offending field, if any.
+    """
+
+    def __init__(self, code: str, message: str, field: str | None = None) -> None:
+        if code not in STATUS:
+            raise ValueError(f"unknown error code {code!r}")
+        super().__init__(message)
+        self.code = code
+        self.message = message
+        self.field = field
+        self.status = STATUS[code]
+
+    def body(self) -> dict[str, Any]:
+        """The JSON error body."""
+        return {"error": {"code": self.code, "message": self.message, "field": self.field}}