QuerryRaspBerry commited on
Commit
8f9bde4
·
1 Parent(s): bc1c646

Mozg odpowiada tylko na podpisane wywolania; bez placu zabaw; awaria to blad, nie pozegnanie

Browse files

- respond(context_json, sig): HMAC-SHA256 z ARIA_BRAIN_SECRET, fail-closed, limity rozmiaru
- _try_it usuniete (publiczne wywolania modelu na kredyt organizacji)
- jezyk po granicach slow, honorowany wybor z Workera; role tylko user/assistant
- blad inferencji rzuca gr.Error zamiast udawac wyjscie ARII
- gradio i huggingface_hub przypiete; README bez hf_oauth

Files changed (4) hide show
  1. README.md +4 -4
  2. app.py +53 -37
  3. brain/orchestrator.py +34 -34
  4. requirements.txt +4 -2
README.md CHANGED
@@ -7,9 +7,6 @@ sdk: gradio
7
  sdk_version: 6.5.1
8
  app_file: app.py
9
  pinned: false
10
- hf_oauth: true
11
- hf_oauth_scopes:
12
- - inference-api
13
  license: apache-2.0
14
  short_description: Adaptive Response & Intelligence Agent
15
  ---
@@ -34,7 +31,10 @@ the brain. A public Space is the wrong place to keep the map of a private backen
34
  and the copies here had drifted behind the live ones anyway.
35
 
36
  ## Run / deploy
37
- - **Brain (this Space):** set `HF_TOKEN` as a Space secret; optionally `ARIA_MODEL`.
 
 
 
38
 
39
  Built on [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub),
40
  and [Cloudflare Workers / Durable Objects / R2](https://developers.cloudflare.com/).
 
7
  sdk_version: 6.5.1
8
  app_file: app.py
9
  pinned: false
 
 
 
10
  license: apache-2.0
11
  short_description: Adaptive Response & Intelligence Agent
12
  ---
 
31
  and the copies here had drifted behind the live ones anyway.
32
 
33
  ## Run / deploy
34
+ - **Brain (this Space):** secrets `HF_TOKEN` (inference) and `ARIA_BRAIN_SECRET` (the
35
+ aria-worker signs every call with it; without it the Space refuses every call);
36
+ optionally `ARIA_MODEL`. There is no public playground: the brain is billed to the
37
+ organisation and answers only signed requests.
38
 
39
  Built on [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub),
40
  and [Cloudflare Workers / Durable Objects / R2](https://developers.cloudflare.com/).
app.py CHANGED
@@ -1,60 +1,76 @@
1
  """ARIA brain — Gradio Space (nextneural-aria.hf.space).
2
 
3
- Exposes one machine API used by the Cloudflare aria-worker:
4
 
5
  api_name="respond"
6
- input : context_json (str) — see brain/orchestrator.respond
7
- output: bubbles_json (str) — JSON array of {bubble, target_event_id, type}
8
 
9
- Also ships a small human tester so the Space is browsable on its own.
 
 
 
10
  """
 
 
 
11
  import json
 
12
 
13
  import gradio as gr
14
 
15
  from brain.orchestrator import respond, MODEL
16
 
 
 
 
17
 
18
- # ---- tiny manual tester: type a line, see ARIA's bubbles ----------------------
19
- def _try_it(text, state):
20
- state = state or {"history": [], "n": 0}
21
- state["n"] += 1
22
- eid = f"evt_{state['n']}"
23
- state["history"].append({"id": eid, "role": "user", "type": "message", "text": text})
24
-
25
- ctx = {
26
- "lang": "auto",
27
- "topic": None,
28
- "turn_events": [eid],
29
- "media": {},
30
- "history": state["history"],
31
- }
32
- out = respond(json.dumps(ctx, ensure_ascii=False))
33
- bubbles = json.loads(out)
34
- for b in bubbles:
35
- state["history"].append({"id": f"a_{state['n']}", "role": "assistant",
36
- "type": "message", "text": b["bubble"]})
37
- pretty = "\n".join(f"• {b['bubble']} [{b['type']}]" for b in bubbles) or "(cisza)"
38
- return pretty, state
39
 
40
 
41
- with gr.Blocks(title="ARIA · brain") as demo:
42
- gr.Markdown(f"## ARIA — brain\nVision + reasoning service. Model: `{MODEL}`")
 
 
 
 
 
 
 
 
 
 
 
 
 
43
 
44
- # Machine API (what the worker calls). Hidden-ish but functional.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45
  with gr.Accordion("Machine API (respond)", open=False):
46
  api_in = gr.Textbox(label="context_json", lines=6)
 
47
  api_out = gr.Textbox(label="bubbles_json", lines=6)
48
- gr.Button("respond").click(respond, inputs=api_in, outputs=api_out,
49
  api_name="respond")
50
 
51
- # Human tester.
52
- gr.Markdown("### Pogadaj z ARIĄ")
53
- st = gr.State()
54
- box = gr.Textbox(label="Ty", placeholder="napisz coś...")
55
- reply = gr.Textbox(label="ARIA", lines=4)
56
- box.submit(_try_it, inputs=[box, st], outputs=[reply, st])
57
-
58
 
59
  if __name__ == "__main__":
60
- demo.queue().launch()
 
1
  """ARIA brain — Gradio Space (nextneural-aria.hf.space).
2
 
3
+ Exposes ONE machine API, used only by the Cloudflare aria-worker:
4
 
5
  api_name="respond"
6
+ inputs : context_json (str), sig (str)
7
+ output : bubbles_json (str) — JSON array of {bubble, target_event_id, type}
8
 
9
+ `sig` is HMAC-SHA256(ARIA_BRAIN_SECRET, context_json) in base64url. Without the
10
+ secret on this Space, or with a bad signature, nothing is sent to the model: the
11
+ brain is paid for by the organisation and must not be callable by anyone who finds
12
+ this URL. The old in-page tester was removed for the same reason.
13
  """
14
+ import base64
15
+ import hashlib
16
+ import hmac
17
  import json
18
+ import os
19
 
20
  import gradio as gr
21
 
22
  from brain.orchestrator import respond, MODEL
23
 
24
+ BRAIN_SECRET = os.environ.get("ARIA_BRAIN_SECRET", "")
25
+ MAX_CONTEXT_CHARS = 60_000
26
+ MAX_HISTORY = 60
27
 
28
+ if not BRAIN_SECRET:
29
+ print("[ARIA] ARIA_BRAIN_SECRET is not set: every call will be refused", flush=True)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
30
 
31
 
32
+ def _b64url_decode(s: str) -> bytes:
33
+ s = (s or "").strip()
34
+ return base64.urlsafe_b64decode(s + "=" * (-len(s) % 4))
35
+
36
+
37
+ def _verify(context_json: str, sig: str) -> bool:
38
+ if not BRAIN_SECRET or not sig or not isinstance(context_json, str):
39
+ return False
40
+ try:
41
+ got = _b64url_decode(sig)
42
+ except Exception:
43
+ return False
44
+ want = hmac.new(BRAIN_SECRET.encode("utf-8"), context_json.encode("utf-8"), hashlib.sha256).digest()
45
+ return hmac.compare_digest(got, want)
46
+
47
 
48
+ def respond_api(context_json: str, sig: str) -> str:
49
+ if not isinstance(context_json, str) or len(context_json) > MAX_CONTEXT_CHARS:
50
+ raise gr.Error("context too large")
51
+ if not _verify(context_json, sig):
52
+ raise gr.Error("unauthorized")
53
+ try:
54
+ ctx = json.loads(context_json)
55
+ except Exception:
56
+ raise gr.Error("bad context")
57
+ if not isinstance(ctx, dict) or len(ctx.get("history") or []) > MAX_HISTORY:
58
+ raise gr.Error("bad context")
59
+ return respond(ctx)
60
+
61
+
62
+ with gr.Blocks(title="ARIA · brain") as demo:
63
+ gr.Markdown(
64
+ f"## ARIA — brain\nVision + reasoning service for app.nnlabs.pl. Model: `{MODEL}`.\n\n"
65
+ "This Space has no public playground: calls must be signed by the NNL backend."
66
+ )
67
  with gr.Accordion("Machine API (respond)", open=False):
68
  api_in = gr.Textbox(label="context_json", lines=6)
69
+ api_sig = gr.Textbox(label="sig", lines=1)
70
  api_out = gr.Textbox(label="bubbles_json", lines=6)
71
+ gr.Button("respond").click(respond_api, inputs=[api_in, api_sig], outputs=api_out,
72
  api_name="respond")
73
 
 
 
 
 
 
 
 
74
 
75
  if __name__ == "__main__":
76
+ demo.queue(default_concurrency_limit=4, max_size=16).launch()
brain/orchestrator.py CHANGED
@@ -13,7 +13,9 @@ We return a JSON string: an array of bubble dicts (see schema.sanitize_bubbles).
13
 
14
  import json
15
  import os
 
16
 
 
17
  from huggingface_hub import InferenceClient
18
 
19
  from .persona import SYSTEM_PROMPT, build_user_frame
@@ -27,36 +29,28 @@ BILL_TO = os.environ.get("ARIA_BILL_TO", "NextNeural") or None
27
  _client = InferenceClient(model=MODEL, token=HF_TOKEN, bill_to=BILL_TO)
28
 
29
 
30
- _LEAVE_PL = ["muszę lecieć", "muszę spadać", "ok to ja lecę", "elo muszę uciekać"]
31
- _BYE_PL = ["siemanko", "nara", "do zobaczenia", "trzymaj się", "pa pa"]
32
-
33
- _LEAVE_EN = ["gotta go", "gotta run", "ok i gotta bounce", "gotta head out"]
34
- _BYE_EN = ["see you", "talk soon", "take care", "bye for now", "cya"]
35
 
36
 
37
  def _lang(ctx: dict) -> str:
 
 
 
 
 
 
38
  users = [h.get("text", "") for h in ctx.get("history", []) if h.get("role") == "user"]
39
  t = (users[-1] if users else "").lower()
40
-
41
  if any(c in t for c in "ąćęłńóśźż"):
42
  return "pl"
43
-
44
- has_pl = any(w in t for w in ["nie", "cześć", "jesteś", "co", "tak", "jak", "siema"])
45
- has_en = any(w in t for w in ["the", "you", "what", "hello", "how", "are", "hey"])
46
-
47
  return "en" if (has_en and not has_pl) else "pl"
48
 
49
 
50
- def _signoff_bubbles(ctx: dict):
51
- import random
52
- leave, bye = (_LEAVE_EN, _BYE_EN) if _lang(ctx) == "en" else (_LEAVE_PL, _BYE_PL)
53
-
54
- return [
55
- {"bubble": random.choice(leave), "target_event_id": None, "type": "proactive"},
56
- {"bubble": random.choice(bye), "target_event_id": None, "type": "proactive"},
57
- ]
58
-
59
-
60
  def _history_to_messages(ctx: dict, sys_extra=""):
61
  """
62
  Role metadata is the ONLY source of truth for who said what.
@@ -71,7 +65,11 @@ def _history_to_messages(ctx: dict, sys_extra=""):
71
 
72
  for h in ctx.get("history", []):
73
  role = h.get("role", "user")
74
- text = h.get("text", "")
 
 
 
 
75
 
76
  # keep raw text ONLY
77
  parts = [{"type": "text", "text": text}]
@@ -85,13 +83,16 @@ def _history_to_messages(ctx: dict, sys_extra=""):
85
  return msgs
86
 
87
 
88
- def respond(context_json: str) -> str:
89
- """Entry called by the Gradio API (api_name='respond')."""
 
90
 
91
- try:
92
- ctx = json.loads(context_json) if isinstance(context_json, str) else (context_json or {})
93
- except Exception:
94
- return "[]"
 
 
95
 
96
  valid_ids = {h.get("id") for h in ctx.get("history", [])}
97
  valid_ids |= set(ctx.get("turn_events", []))
@@ -117,14 +118,13 @@ def respond(context_json: str) -> str:
117
  temperature=0.85,
118
  top_p=0.95,
119
  )
120
-
121
  raw = completion.choices[0].message.content or "[]"
122
-
123
  except Exception as e:
124
- import traceback, sys
125
- print("ARIA inference error:", repr(e), file=sys.stderr)
126
- traceback.print_exc()
127
- return json.dumps(_signoff_bubbles(ctx), ensure_ascii=False)
 
128
 
129
  bubbles = sanitize_bubbles(raw, valid_ids)
130
- return json.dumps(bubbles, ensure_ascii=False)
 
13
 
14
  import json
15
  import os
16
+ import re
17
 
18
+ import gradio as gr
19
  from huggingface_hub import InferenceClient
20
 
21
  from .persona import SYSTEM_PROMPT, build_user_frame
 
29
  _client = InferenceClient(model=MODEL, token=HF_TOKEN, bill_to=BILL_TO)
30
 
31
 
32
+ _PL_WORDS = ("nie", "cześć", "czesc", "jesteś", "jestes", "co", "tak", "jak", "siema", "dzięki", "dzieki",
33
+ "jestem", "mam", "robisz", "elo", "dobra", "spoko")
34
+ _EN_WORDS = ("the", "you", "what", "hello", "how", "are", "hey", "hi", "yes", "yeah", "thanks", "im", "me")
 
 
35
 
36
 
37
  def _lang(ctx: dict) -> str:
38
+ """Polish-first. The worker may pass a decision ("pl"/"en"); otherwise look at
39
+ the user's latest message with WORD boundaries (``"tak" in "take"`` used to make
40
+ "take care" Polish)."""
41
+ hinted = ctx.get("lang")
42
+ if hinted in ("pl", "en"):
43
+ return hinted
44
  users = [h.get("text", "") for h in ctx.get("history", []) if h.get("role") == "user"]
45
  t = (users[-1] if users else "").lower()
 
46
  if any(c in t for c in "ąćęłńóśźż"):
47
  return "pl"
48
+ words = set(re.findall(r"[a-ząćęłńóśźż']+", t))
49
+ has_pl = any(w in words for w in _PL_WORDS)
50
+ has_en = any(w in words for w in _EN_WORDS)
 
51
  return "en" if (has_en and not has_pl) else "pl"
52
 
53
 
 
 
 
 
 
 
 
 
 
 
54
  def _history_to_messages(ctx: dict, sys_extra=""):
55
  """
56
  Role metadata is the ONLY source of truth for who said what.
 
65
 
66
  for h in ctx.get("history", []):
67
  role = h.get("role", "user")
68
+ if role == "system" and h.get("type") == "system_idle":
69
+ role = "user" # the idle marker reads as an observation, never as an instruction
70
+ if role not in ("user", "assistant"):
71
+ role = "user"
72
+ text = str(h.get("text", ""))[:4000]
73
 
74
  # keep raw text ONLY
75
  parts = [{"type": "text", "text": text}]
 
83
  return msgs
84
 
85
 
86
+ def respond(ctx) -> str:
87
+ """Entry called from app.py after signature verification. `ctx` is the parsed
88
+ context dict (a JSON string is still accepted for local testing)."""
89
 
90
+ if isinstance(ctx, str):
91
+ try:
92
+ ctx = json.loads(ctx)
93
+ except Exception:
94
+ return "[]"
95
+ ctx = ctx or {}
96
 
97
  valid_ids = {h.get("id") for h in ctx.get("history", [])}
98
  valid_ids |= set(ctx.get("turn_events", []))
 
118
  temperature=0.85,
119
  top_p=0.95,
120
  )
 
121
  raw = completion.choices[0].message.content or "[]"
 
122
  except Exception as e:
123
+ # No in-character goodbye here: the worker turns a failed call into a visible
124
+ # error line, so an outage never looks like ARIA leaving.
125
+ import sys
126
+ print("ARIA inference error:", type(e).__name__, file=sys.stderr)
127
+ raise gr.Error("model unavailable")
128
 
129
  bubbles = sanitize_bubbles(raw, valid_ids)
130
+ return json.dumps(bubbles, ensure_ascii=False)
requirements.txt CHANGED
@@ -1,2 +1,4 @@
1
- gradio>=6.5.1
2
- huggingface_hub>=0.29.0
 
 
 
1
+ # Pinned so a Space restart cannot pull a different Gradio and change the API the
2
+ # worker depends on. gradio matches sdk_version in README.md.
3
+ gradio==6.5.1
4
+ huggingface_hub==1.15.0