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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
|
#!/usr/bin/env python3
"""Hermetic end-to-end test of the ffwd-server HTTP API.
Builds the tiny synthetic fixture (the gen_fixture test helper, the same model
the C tests write) and the ffwd-server binary, launches the server on a free
loopback port against the fixture across standard, contextual, and late/rerank
model slots, then drives the full HTTP request path and asserts on it: routing,
auth, malformed input, validation 422s, output encodings, Matryoshka
truncation, text_type prompting, micro-batching, exact token counts, and rerank
ordering.
The fixture weights are deterministic but meaningless, so this checks PROTOCOL
and STRUCTURE, not embedding quality. It is the out-of-process successor to the
in-process C test that used to live in tests/test_server.c; the server's
internal helpers are unit-tested separately under tools/server.
Usage: tests/check_server_api.py [--server PATH] [--gen PATH] [-v]
"""
from __future__ import annotations
import argparse
import json
import os
import socket
import subprocess
import sys
import tempfile
import time
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
HOST = "127.0.0.1"
API_KEY = "tt-test-key"
STD_DIM = 1024
GTE_DIM = 1536
LATE_TOKEN_DIM = 128
# The retrieval task text. Qwen3 has no trailing space after "Query:"; GTE-Qwen2
# adds one. The instruction holds a newline and colons, so the model-spec parser
# must take query_instruct= as the final key (it does) and unescape \n.
QWEN_INSTRUCT = (
"Instruct: Given a web search query, retrieve relevant passages "
"that answer the query\nQuery:"
)
GTE_INSTRUCT = (
"Instruct: Given a web search query, retrieve relevant passages "
"that answer the query\nQuery: "
)
PORT = 0
VERBOSE = False
failures = 0
def fail(what: str) -> None:
global failures
failures += 1
print(f"FAIL: {what}", file=sys.stderr)
def expect(cond: bool, what: str) -> None:
if not cond:
fail(what)
# --------------------------------------------------------------------------
# Build + fixture + launch
# --------------------------------------------------------------------------
def ensure_tool(path: str, make_dir: str, make_target: str) -> None:
"""Build a Make target if its output binary is missing."""
if not os.path.exists(path):
subprocess.run(["make", "-C", make_dir, make_target], cwd=ROOT,
check=True, stdout=subprocess.DEVNULL)
if not os.path.exists(path):
sys.exit(f"required binary not found and could not be built: {path}")
def write_fixtures(gen: str, root: str) -> dict[str, str]:
"""Write the four fixture model directories and return their paths.
std-main and ctx-main share one base model; the late model adds a Dense
projection; qwen3 and gte-qwen2 differ in architecture and hidden size.
"""
dirs = {
"base": os.path.join(root, "base"),
"late": os.path.join(root, "late"),
"qwen": os.path.join(root, "qwen"),
"gte": os.path.join(root, "gte"),
}
specs = [
(dirs["base"], ["--model", "base", "--hidden", str(STD_DIM)]),
(dirs["late"], ["--model", "base", "--hidden", str(STD_DIM),
"--late-dim", str(LATE_TOKEN_DIM)]),
(dirs["qwen"], ["--model", "qwen3", "--hidden", str(STD_DIM)]),
(dirs["gte"], ["--model", "qwen2", "--hidden", str(GTE_DIM)]),
]
for out, args in specs:
os.mkdir(out)
subprocess.run([gen, "--out", out, *args], check=True,
stdout=subprocess.DEVNULL)
return dirs
def free_port() -> int:
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.bind((HOST, 0))
port = s.getsockname()[1]
s.close()
return port
def cli_instruct(text: str) -> str:
"""Encode a newline as the \\n the model-spec parser unescapes."""
return text.replace("\n", "\\n")
def server_argv(server: str, dirs: dict[str, str], port: int) -> list[str]:
return [
server,
"--host", HOST,
"--port", str(port),
"--api-key", API_KEY,
"-b", "2",
"--batch-wait-us", "1000",
"--model", f"std-main={dirs['base']}:api=perplexity:min_dim=128",
"--contextual-model", f"ctx-main={dirs['base']}:api=perplexity:min_dim=128",
"--late-model", f"late-main={dirs['late']}:api=perplexity:min_dim=128",
"--model", f"qwen-search={dirs['qwen']}:min_dim=32"
f":query_instruct={cli_instruct(QWEN_INSTRUCT)}",
"--model", f"gte-search={dirs['gte']}:min_dim=1536"
f":query_instruct={cli_instruct(GTE_INSTRUCT)}",
]
def wait_ready(proc: subprocess.Popen, log_path: str, timeout_s: float = 40.0) -> None:
"""Poll the server until a real embeddings request returns 200.
The model load happens in the worker thread, so the first request may block
or be refused while the server comes up; retry until it succeeds.
"""
deadline = time.monotonic() + timeout_s
warmup = {"model": "std-main", "input": "ready?", "encoding_format": "float"}
while time.monotonic() < deadline:
if proc.poll() is not None:
_die_with_log(proc, log_path, "server exited during startup")
try:
status, _ = request("POST", "/v1/embeddings", body=warmup, timeout=5)
if status == 200:
return
except OSError:
pass
time.sleep(0.1)
_die_with_log(proc, log_path, "server did not become ready in time")
def _die_with_log(proc: subprocess.Popen, log_path: str, msg: str) -> None:
try:
with open(log_path) as f:
tail = "".join(f.readlines()[-20:])
except OSError:
tail = "(no server log)"
try:
proc.kill()
except OSError:
pass
sys.exit(f"{msg}\n--- server log tail ---\n{tail}")
# --------------------------------------------------------------------------
# HTTP
# --------------------------------------------------------------------------
def request(method: str, path: str, body=None, auth: str | None = API_KEY,
timeout: float = 30.0) -> tuple[int, str]:
"""One request over a fresh connection; returns (status, body text).
body: dict/list -> JSON; str -> sent verbatim (for malformed-JSON checks);
None -> no body.
"""
if isinstance(body, (dict, list)):
data = json.dumps(body).encode()
elif isinstance(body, str):
data = body.encode()
else:
data = None
headers = {}
if data is not None:
headers["Content-Type"] = "application/json"
if auth:
headers["Authorization"] = f"Bearer {auth}"
import http.client
conn = http.client.HTTPConnection(HOST, PORT, timeout=timeout)
try:
conn.request(method, path, body=data, headers=headers)
resp = conn.getresponse()
text = resp.read().decode("utf-8", "replace")
if VERBOSE:
print(f" {method} {path} -> {resp.status}", file=sys.stderr)
return resp.status, text
finally:
conn.close()
def request_raw(raw: bytes, timeout: float = 10.0) -> int:
"""Send raw bytes on the wire and return the parsed status code (or -1)."""
with socket.create_connection((HOST, PORT), timeout=timeout) as s:
s.sendall(raw)
s.settimeout(timeout)
data = b""
try:
while b"\r\n" not in data:
chunk = s.recv(4096)
if not chunk:
break
data += chunk
except OSError:
pass
try:
return int(data.split(b"\r\n", 1)[0].split()[1])
except (IndexError, ValueError):
return -1
def embeddings(model: str, **fields) -> tuple[int, str]:
return request("POST", "/v1/embeddings", body={"model": model, **fields})
def max_absdiff(a, b) -> float:
if not isinstance(a, list) or not isinstance(b, list) or len(a) != len(b):
return float("inf")
return max((abs(x - y) for x, y in zip(a, b)), default=0.0)
# --------------------------------------------------------------------------
# Test groups (ported from tests/test_server.c)
# --------------------------------------------------------------------------
def test_routing_and_auth() -> None:
expect(request("OPTIONS", "/v1/embeddings", auth=None)[0] == 204,
"OPTIONS preflight returns 204")
expect(request("GET", "/v1/embeddings")[0] == 404, "GET on embeddings is 404")
expect(request("POST", "/nope", body={})[0] == 404, "unknown path is 404")
ok = {"model": "std-main", "input": "hello world", "encoding_format": "float"}
expect(request("POST", "/v1/embeddings", body=ok, auth=None)[0] == 401,
"missing auth is 401")
expect(request("POST", "/v1/embeddings", body=ok, auth="wrong-key")[0] == 401,
"wrong key is 401")
expect(request_raw(b"GARBAGE\r\n\r\n") == 400, "malformed HTTP is 400")
expect(request("POST", "/v1/embeddings", body="{nope")[0] == 422,
"malformed JSON is 422")
def test_validation() -> None:
cases = [
({"input": "hi"}, "missing model"),
({"model": "foo", "input": "hi"}, "unknown model"),
({"model": "ctx-main", "input": "hi"}, "contextual model on /v1/embeddings"),
({"model": "std-missing", "input": "hi"}, "unloaded model"),
({"model": "std-main", "input": "hi", "encoding_format": "yaml"},
"bad encoding enum"),
({"model": "std-main", "input": "hi", "dimensions": 64},
"dimensions below min_dim"),
({"model": "std-main", "input": []}, "empty input array"),
]
for body, label in cases:
expect(request("POST", "/v1/embeddings", body=body)[0] == 422,
f"422 for {label}")
expect(request("POST", "/v1/contextualizedembeddings",
body={"model": "std-main", "input": [["a"]]})[0] == 422,
"standard model on contextual endpoint is 422")
expect(request("POST", "/v1/contextualizedembeddings",
body={"model": "ctx-missing", "input": [["a"]]})[0] == 422,
"unknown contextual model is 422")
def test_standard_embeddings() -> None:
status, text = embeddings("std-main", input="hello world", encoding_format="float")
expect(status == 200, "standard float request is 200")
if status == 200:
root = json.loads(text)
data = root["data"]
expect(len(data) == 1, "one embedding for one input")
emb = data[0]["embedding"]
expect(isinstance(emb, list) and len(emb) == STD_DIM,
"float embedding has full dimension")
expect(data[0]["index"] == 0, "index is 0")
expect(all(isinstance(v, (int, float)) for v in emb),
"embedding values are numbers")
expect(any(v != 0.0 for v in emb), "embedding is not all zeros")
expect(root["usage"]["total_tokens"] >= 1, "usage reports tokens")
# Batch: one row per input, ascending indexes, distinct vectors.
status, text = embeddings("std-main", input=["hello", "world", "held"],
encoding_format="float")
expect(status == 200, "batch request is 200")
if status == 200:
data = json.loads(text)["data"]
expect(len(data) == 3, "three embeddings for three inputs")
expect([d["index"] for d in data] == [0, 1, 2], "indexes are 0..n-1")
expect(max_absdiff(data[0]["embedding"], data[1]["embedding"]) > 0.0,
"different inputs give different vectors")
# Perplexity default encoding is base64_int8: a string of the int8 base64.
b64_len = ((STD_DIM + 2) // 3) * 4
for label, fields in [("default", {}),
("explicit", {"encoding_format": "base64_int8"})]:
status, text = embeddings("std-main", input="hello", **fields)
expect(status == 200, f"{label} int8 request is 200")
if status == 200:
emb = json.loads(text)["data"][0]["embedding"]
expect(isinstance(emb, str) and len(emb) == b64_len,
f"{label} encoding is base64_int8 string")
expect(embeddings("std-main", input="hello", encoding_format="base64")[0] == 422,
"OpenAI base64 is rejected on a Perplexity model")
# Matryoshka truncation via dimensions.
status, text = embeddings("std-main", input="hello", dimensions=128,
encoding_format="float")
expect(status == 200, "matryoshka request is 200")
if status == 200:
emb = json.loads(text)["data"][0]["embedding"]
expect(isinstance(emb, list) and len(emb) == 128, "truncated to 128 dims")
def test_qwen_mrl_and_text_type() -> None:
status, text = embeddings("qwen-search", input="hello", encoding_format="float")
expect(status == 200, "qwen full request is 200")
full = json.loads(text)["data"][0]["embedding"] if status == 200 else None
status, text = embeddings("qwen-search", input="hello", dimensions=32,
encoding_format="float")
expect(status == 200, "qwen MRL-32 request is 200")
if status == 200:
root = json.loads(text)
mrl = root["data"][0]["embedding"]
expect(isinstance(full, list) and len(full) == STD_DIM, "qwen full is 1024")
expect(isinstance(mrl, list) and len(mrl) == 32, "qwen MRL is 32")
if isinstance(full, list) and isinstance(mrl, list):
expect(max((abs(full[i] - mrl[i]) for i in range(32)), default=0.0) > 0.0,
"MRL re-normalizes the truncated prefix")
# hello plus Qwen3's terminal token = 2 tokens.
expect(root["usage"]["total_tokens"] == 2, "qwen usage counts the EOT token")
# text_type=query must equal manually prepending the instruction with
# text_type=document: same tokens, hence the same vector.
for model, instruct, dim in [("qwen-search", QWEN_INSTRUCT, STD_DIM),
("gte-search", GTE_INSTRUCT, GTE_DIM)]:
sa, ta = embeddings(model, input="hello", text_type="query",
encoding_format="float")
sb, tb = embeddings(model, input=instruct + "hello", text_type="document",
encoding_format="float")
expect(sa == 200 and sb == 200, f"{model} text_type requests are 200")
if sa == 200 and sb == 200:
qe = json.loads(ta)["data"][0]["embedding"]
me = json.loads(tb)["data"][0]["embedding"]
expect(isinstance(qe, list) and len(qe) == dim, f"{model} query dim {dim}")
expect(max_absdiff(qe, me) < 1e-6,
f"{model} text_type=query equals the manual instruction")
expect(embeddings("std-main", input="hello", text_type="query")[0] == 422,
"text_type on a model without a query instruction is 422")
expect(embeddings("qwen-search", input="hello", text_type="passage")[0] == 422,
"invalid text_type value is 422")
def test_contextual() -> None:
expect(request("POST", "/v1/contextualizedembeddings",
body={"model": "ctx-main", "input": "hi"})[0] == 422,
"contextual input must be an array of chunk arrays")
expect(request("POST", "/v1/contextualizedembeddings",
body={"model": "ctx-main", "input": [[""]]})[0] == 422,
"contextual chunks must be non-empty")
status, text = request("POST", "/v1/contextualizedembeddings",
body={"model": "ctx-main",
"input": [["hello", "world"], ["held"]],
"encoding_format": "float"})
expect(status == 200, "contextual happy path is 200")
multi = json.loads(text) if status == 200 else None
if multi:
docs = multi["data"]
expect(len(docs) == 2, "two documents")
for di, doc in enumerate(docs):
expect(doc["index"] == di, f"doc {di} index")
chunks = doc["data"]
expect(len(chunks) == (2 if di == 0 else 1), f"doc {di} chunk count")
for ci, chunk in enumerate(chunks):
expect(chunk["index"] == ci, f"doc {di} chunk {ci} index")
expect(isinstance(chunk["embedding"], list)
and len(chunk["embedding"]) == STD_DIM,
f"doc {di} chunk {ci} embedding dim")
# hello=1, world=3, held=2, one separator in doc 0 -> 7.
expect(multi["usage"]["total_tokens"] == 7, "contextual token accounting")
# A single-chunk document covers the whole sequence, so it must match the
# standard embedding of the same text; a chunk beside a neighbor must not.
status, text = request("POST", "/v1/contextualizedembeddings",
body={"model": "ctx-main", "input": [["hello"]],
"encoding_format": "float"})
single = json.loads(text) if status == 200 else None
status, text = embeddings("std-main", input="hello", encoding_format="float")
std = json.loads(text) if status == 200 else None
if multi and single and std:
std_emb = std["data"][0]["embedding"]
alone = single["data"][0]["data"][0]["embedding"]
in_doc = multi["data"][0]["data"][0]["embedding"]
expect(max_absdiff(std_emb, alone) < 1e-4,
"single-chunk contextual equals standard embedding")
expect(max_absdiff(alone, in_doc) > 1e-3,
"whole-document attention changes a chunk beside a neighbor")
def test_rerank() -> None:
expect(request("POST", "/v1/rerank",
body={"model": "std-main", "query": "hello",
"documents": ["world"]})[0] == 422,
"rerank on a standard model is 422")
expect(request("POST", "/v1/rerank",
body={"model": "late-main", "query": "",
"documents": ["world"]})[0] == 422,
"empty query is 422")
expect(request("POST", "/v1/rerank",
body={"model": "late-main", "query": "hello",
"documents": []})[0] == 422,
"empty documents is 422")
expect(request("POST", "/v1/rerank",
body={"model": "late-main", "query": "hello",
"documents": ["world"], "top_n": 1, "top_k": 1})[0] == 422,
"top_n and top_k together is 422")
status, text = request("POST", "/v1/rerank",
body={"model": "late-main", "query": "hello",
"documents": ["hello!", "world",
'quoted "document"'],
"top_n": 2, "return_documents": True})
expect(status == 200, "rerank happy path is 200")
if status == 200:
root = json.loads(text)
expect(root.get("object") == "list", "rerank object is list")
expect(root.get("model") == "late-main", "rerank echoes the model")
results = root["results"]
expect(len(results) == 2, "top_n=2 returns two results")
scores = [r["relevance_score"] for r in results]
expect(scores == sorted(scores, reverse=True), "results sorted by score")
idxs = [r["index"] for r in results]
expect(all(0 <= i < 3 for i in idxs) and len(set(idxs)) == 2,
"result indexes are distinct and in range")
expect(all(isinstance(r.get("document"), str) for r in results),
"return_documents includes the document text")
usage = root["usage"]
expect(usage["query_tokens"] == 32, "query expands to 32 tokens")
expect(usage["document_tokens"] > 0, "documents contribute tokens")
expect(usage["total_tokens"] == usage["query_tokens"] + usage["document_tokens"],
"rerank token accounting is consistent")
status, text = request("POST", "/v1/rerank",
body={"model": "late-main", "query": "hello",
"documents": ["hello", "world"], "top_k": 1})
expect(status == 200, "rerank top_k alias is 200")
if status == 200:
results = json.loads(text)["results"]
expect(len(results) == 1, "top_k=1 returns one result")
expect("document" not in results[0],
"documents are omitted unless requested")
def test_concurrent() -> None:
"""More in-flight requests than the micro-batch cap (-b 2)."""
import threading
n = 6
results = [None] * n
body = {"model": "std-main", "input": "hello world", "encoding_format": "float"}
def worker(i: int) -> None:
try:
results[i] = request("POST", "/v1/embeddings", body=body)[0]
except OSError as e: # noqa: BLE001
results[i] = f"error: {e}"
threads = [threading.Thread(target=worker, args=(i,)) for i in range(n)]
for t in threads:
t.start()
for t in threads:
t.join()
expect(all(r == 200 for r in results),
f"all {n} concurrent requests return 200 (got {results})")
def test_keepalive() -> None:
"""One HTTP/1.1 connection is reused across requests (no Connection: close)."""
import http.client
conn = http.client.HTTPConnection(HOST, PORT, timeout=30)
body = json.dumps({"model": "std-main", "input": ["hello", "world"],
"encoding_format": "float"}).encode()
headers = {"Content-Type": "application/json",
"Authorization": f"Bearer {API_KEY}"}
first_sock = None
try:
for i in range(2):
conn.request("POST", "/v1/embeddings", body=body, headers=headers)
resp = conn.getresponse()
resp.read()
expect(resp.status == 200, f"keep-alive request {i} is 200")
expect((resp.getheader("Connection") or "").lower() != "close",
"server does not send Connection: close")
expect(conn.sock is not None, "connection stays open after response")
if i == 0:
first_sock = conn.sock
else:
expect(conn.sock is first_sock, "connection is reused, not reconnected")
finally:
conn.close()
# --------------------------------------------------------------------------
def main() -> int:
global PORT, VERBOSE
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("--server", default=os.path.join(ROOT, "tools", "server",
"ffwd-server"),
help="path to the ffwd-server binary")
ap.add_argument("--gen", default=os.path.join(ROOT, "tests", "gen_fixture"),
help="path to the gen_fixture helper")
ap.add_argument("-v", "--verbose", action="store_true")
args = ap.parse_args()
VERBOSE = args.verbose
ensure_tool(args.server, "tools/server", "all")
ensure_tool(args.gen, "tests", "gen_fixture")
with tempfile.TemporaryDirectory(prefix="ffwd-server-check-") as workdir:
dirs = write_fixtures(args.gen, workdir)
PORT = free_port()
log_path = os.path.join(workdir, "server.log")
with open(log_path, "w") as log:
proc = subprocess.Popen(server_argv(args.server, dirs, PORT),
stdout=log, stderr=subprocess.STDOUT)
try:
wait_ready(proc, log_path)
test_routing_and_auth()
test_validation()
test_standard_embeddings()
test_keepalive()
test_qwen_mrl_and_text_type()
test_contextual()
test_rerank()
test_concurrent()
# Graceful shutdown: SIGTERM must unwind the server to a clean exit
# code 0 (the old C test's `ctx.rc == 0` check).
proc.terminate()
try:
expect(proc.wait(timeout=10) == 0, "server exits cleanly on SIGTERM")
except subprocess.TimeoutExpired:
fail("server did not exit within 10s of SIGTERM")
finally:
if proc.poll() is None:
proc.kill()
proc.wait()
if failures:
print(f"{failures} server API check(s) failed", file=sys.stderr)
return 1
print("ok: ffwd-server HTTP request-path checks passed")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|