File size: 10,205 Bytes
466cc3d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6d9c72b
 
466cc3d
 
6d9c72b
466cc3d
 
6d9c72b
cf0a8ed
466cc3d
 
cf0a8ed
 
6d9c72b
 
 
466cc3d
 
6d9c72b
466cc3d
6d9c72b
cf0a8ed
466cc3d
6d9c72b
 
 
 
466cc3d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6d9c72b
 
466cc3d
 
447af01
 
 
466cc3d
447af01
 
466cc3d
 
447af01
 
466cc3d
 
447af01
 
466cc3d
 
447af01
 
466cc3d
 
447af01
6d9c72b
466cc3d
 
 
6d9c72b
466cc3d
 
 
 
 
 
 
 
6d9c72b
 
466cc3d
 
 
6d9c72b
466cc3d
 
 
 
 
 
 
 
 
 
 
 
 
 
6d9c72b
466cc3d
 
 
 
6d9c72b
 
 
466cc3d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6d9c72b
466cc3d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6d9c72b
 
 
466cc3d
 
 
447af01
466cc3d
 
 
 
 
 
 
 
 
 
 
 
 
 
6d9c72b
 
466cc3d
 
 
6d9c72b
 
466cc3d
6d9c72b
 
 
 
 
 
 
 
466cc3d
 
 
6d9c72b
466cc3d
6d9c72b
 
 
466cc3d
6d9c72b
466cc3d
 
 
 
6d9c72b
466cc3d
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
"""FastAPI MCP-compatible server exposing ContextForge tools.

The server uses a FastAPI lifespan to construct the heavy components once
(`ContextRegistry`, `ContextCompressor`, `CompressionCoordinator`,
`MetricsCollector`, `VLLMClient`) and stores them on `app.state`. Endpoints
read these via the dependency-getter functions defined below; tests
override the same getters via `app.dependency_overrides` so endpoint logic
runs against fakes without ever entering the lifespan.

Important contracts:
- /health returns the metrics-supplied GPU label, never the request body.
- Endpoints log only metadata (agent_id, lengths) — never the raw context —
  so request payloads cannot leak via stdout/stderr.
"""
from __future__ import annotations

import asyncio
import logging
from contextlib import asynccontextmanager
from typing import Any, AsyncIterator

from fastapi import Depends, FastAPI, Request
from fastapi.responses import JSONResponse

from apohara_context_forge.config import settings
from apohara_context_forge.compression.compressor import ContextCompressor
from apohara_context_forge.compression.coordinator import CompressionCoordinator
from apohara_context_forge.metrics.collector import MetricsCollector
from apohara_context_forge.models import (
    CompressionDecision,
    ContextEntry,
    ContextMatch,
    ContextRegistration,
    Degradation,
    MetricsSnapshot,
    OptimizedContextRequest,
)
from apohara_context_forge.registry.context_registry import ContextRegistry
from apohara_context_forge.serving.vllm_client import VLLMClient

logger = logging.getLogger(__name__)


# ---------------------------------------------------------------------------
# Lifespan — constructs heavy components once and tears them down on shutdown.
# ---------------------------------------------------------------------------

@asynccontextmanager
async def lifespan(app: FastAPI) -> AsyncIterator[None]:
    """Build app.state.* once; release resources on shutdown.

    Tests bypass the production heavy path either by NOT entering the
    `with TestClient(app) as client:` context (so this lifespan never fires)
    or by monkeypatching the constructor classes referenced by name on this
    module before entering the context.
    """
    app.state.registry = ContextRegistry()
    app.state.compressor = ContextCompressor()
    app.state.coordinator = CompressionCoordinator(
        registry=app.state.registry,
        compressor=app.state.compressor,
    )
    app.state.metrics = MetricsCollector()
    app.state.vllm = VLLMClient()
    logger.info(
        "ContextForge started on %s:%s (vLLM %s, model %s)",
        settings.contextforge_host,
        settings.contextforge_port,
        settings.vllm_base_url,
        settings.vllm_model,
    )
    try:
        yield
    finally:
        # Best-effort teardown — never let cleanup errors mask the original
        # request error during shutdown.
        clear = getattr(app.state.registry, "clear", None)
        if clear is not None:
            try:
                await clear()
            except Exception as exc:
                logger.warning("registry.clear() failed: %s", exc)
        aclose = getattr(app.state.vllm, "aclose", None)
        if aclose is not None:
            try:
                await aclose()
            except Exception as exc:
                logger.warning("vllm.aclose() failed: %s", exc)


app = FastAPI(title="ContextForge", version="0.1.0", lifespan=lifespan)


# Module-level globals kept for callers that import the server outside a
# lifespan-managed TestClient (e.g. ad-hoc REPL probes). Endpoints prefer
# `request.app.state.*` via the dependency getters below.
registry = ContextRegistry()
metrics = MetricsCollector()
compressor: ContextCompressor | None = None
coordinator: CompressionCoordinator | None = None


# ---------------------------------------------------------------------------
# Dependency getters — keys for app.dependency_overrides in tests.
# ---------------------------------------------------------------------------

def get_registry(request: Request) -> ContextRegistry:
    return getattr(request.app.state, "registry", registry)


def get_metrics(request: Request) -> MetricsCollector:
    return getattr(request.app.state, "metrics", metrics)


def get_compressor(request: Request) -> Any:
    return getattr(request.app.state, "compressor", compressor)


def get_coordinator(request: Request) -> Any:
    return getattr(request.app.state, "coordinator", coordinator)


# ---------------------------------------------------------------------------
# /health — never raises. Reports {"status": "ok"|"degraded", "gpu": <label>}.
# ---------------------------------------------------------------------------

@app.get("/health")
async def health_check(metrics: MetricsCollector = Depends(get_metrics)) -> dict:
    try:
        label = metrics._resolve_gpu_label()
        return {"status": "ok", "gpu": label}
    except Exception:
        # Anything failing here is a soft-degrade — clients keep polling.
        return {"status": "degraded", "gpu": "unknown"}


# ---------------------------------------------------------------------------
# /tools/register_context
# ---------------------------------------------------------------------------

@app.post("/tools/register_context", response_model=ContextEntry)
async def register_context(
    registration: ContextRegistration,
    registry: ContextRegistry = Depends(get_registry),
    metrics: MetricsCollector = Depends(get_metrics),
) -> ContextEntry:
    """Register an agent's context. Strict body validation: missing field,
    empty agent_id, or extra fields all yield 422 (handled by Pydantic)."""
    # Log metadata only — NEVER the raw context (sentinel-leakage test).
    logger.info(
        "register_context agent_id=%s ctx_len=%d",
        registration.agent_id,
        len(registration.context),
    )
    entry = await registry.register(registration.agent_id, registration.context)
    # The simple register endpoint does not run cross-agent dedup, so we
    # always report `matched=False`. The richer pipeline path uses
    # registry.register_agent and reports its own match telemetry.
    metrics.record_register(False)
    return entry


# ---------------------------------------------------------------------------
# /tools/get_optimized_context
# ---------------------------------------------------------------------------

def _passthrough_decision(context: str) -> CompressionDecision:
    """Build the safe fallback returned with HTTP 503 when the coordinator
    raises. Callers receive a structured payload and can re-issue or fall
    back to the original context themselves."""
    return CompressionDecision(
        strategy="passthrough",
        final_context=context,
        compressed_context=context,
        shared_prefix="",
        original_tokens=0,
        final_tokens=0,
        tokens_saved=0,
        rationale="coordinator_unavailable",
        savings_pct=0.0,
    )


@app.post("/tools/get_optimized_context")
async def get_optimized_context(
    request: OptimizedContextRequest,
    coordinator: Any = Depends(get_coordinator),
    metrics: MetricsCollector = Depends(get_metrics),
):
    """Return a compression decision. On coordinator failure return 503 with
    a passthrough decision body — the client gets a structured response, not
    a 500 stack trace, and metrics.record_decision is NOT called."""
    logger.info(
        "get_optimized_context agent_id=%s ctx_len=%d",
        request.agent_id,
        len(request.context),
    )
    try:
        decision = await coordinator.decide(request.agent_id, request.context)
    except Exception as exc:
        # Don't log the body — only the error class. The sentinel-leakage
        # test asserts no log record contains the original context string.
        logger.warning(
            "coordinator.decide failed for agent_id=%s: %s",
            request.agent_id,
            type(exc).__name__,
        )
        fallback = _passthrough_decision(request.context)
        return JSONResponse(status_code=503, content=fallback.model_dump(mode="json"))

    metrics.record_decision(decision)
    return decision


# ---------------------------------------------------------------------------
# /metrics/snapshot
# ---------------------------------------------------------------------------

@app.get("/metrics/snapshot", response_model=MetricsSnapshot)
async def metrics_snapshot_endpoint(
    metrics: MetricsCollector = Depends(get_metrics),
    compressor: Any = Depends(get_compressor),
) -> MetricsSnapshot:
    """Aggregate snapshot. We pull `current_model` and `degradations` from the
    compressor (which the lifespan owns) and forward them to the collector,
    which doesn't itself know about compressor identity."""
    current_model = getattr(compressor, "current_model", None) or "xlm-roberta-large"
    degradations = list(getattr(compressor, "degradations", []) or [])
    return await metrics.snapshot(
        current_compressor_model=current_model,
        compressor_degradations=degradations,
    )


# ---------------------------------------------------------------------------
# Root
# ---------------------------------------------------------------------------

@app.get("/")
async def root() -> dict:
    return {
        "service": "ContextForge",
        "version": "0.1.0",
        "description": "The shared context compiler for multi-agent LLM systems",
        "docs": "/docs",
    }


# ---------------------------------------------------------------------------
# Background metrics loop — opt-in helper for production runs.
# ---------------------------------------------------------------------------

async def metrics_loop() -> None:
    while True:
        try:
            await asyncio.sleep(30)
            snap = await metrics.snapshot()
            logger.info(
                "Metrics: VRAM=%.1fGB TTFT=%.1fms Dedup=%.1f%%",
                snap.vram_used_gb,
                snap.ttft_ms,
                snap.dedup_rate,
            )
        except Exception as exc:
            logger.error("Metrics collection error: %s", exc)