Jibo-Revival-Group/JiboOs
8
1
Fork 1
Code Issues 1 Pull Requests Actions Packages Projects Releases 1 Wiki Activity

umm ai bridge upgrades

Browse Source
This commit is contained in:
Kevin
2026-04-01 20:32:01 +03:00
parent f6e5f2db36
commit aaf621460d
5 changed files with 474 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

169
V3.1/build/ai_bridge_server/server.py
View File

@@ -1,15 +1,23 @@
#!/usr/bin/env python3
"""Minimal AI Bridge server for Jibo.
"""Streaming AI Bridge server for Jibo.
Replicates the original Jibo server pattern: responds sentence-by-sentence
so the robot can start speaking immediately while the LLM is still generating.
Endpoints:
- POST /v1/chat/text {"text": "..."} -> {"reply": "..."}
- POST /v1/chat/audio {"wav_base64": "...", "sample_rate": 16000} -> {"reply": "...", "text": "<transcript>"}
- POST /v1/chat/text {"text": "..."} -> chunked NDJSON, one line per sentence:
{"sentence": "First sentence.", "done": false}
{"sentence": "Second sentence!", "done": false}
{"sentence": "", "done": true, "reply": "First sentence. Second sentence!"}
Legacy (non-streaming) clients still work: the final line has the full "reply".
- POST /v1/chat/audio {"wav_base64": "...", "sample_rate": 16000} -> same streaming format
LLM:
- Uses Ollama Chat API by default: http://localhost:11434/api/chat
- Uses Ollama Chat API with streaming: http://localhost:11434/api/chat
Env:
OLLAMA_URL (default: http://127.0.0.1:11434/api/chat)
OLLAMA_MODEL (default: phi3.5)
OLLAMA_MODEL (default: smollm2:135m)
STT (optional, for /audio):
- If `faster-whisper` is installed, it will be used.
@@ -63,8 +71,9 @@ def _read_json(handler: BaseHTTPRequestHandler) -> dict:
def _ollama_chat(user_text: str) -> str:
"""Non-streaming fallback (used by /audio endpoint)."""
ollama_url = os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434/api/chat")
model = os.environ.get("OLLAMA_MODEL", "phi3.5")
model = os.environ.get("OLLAMA_MODEL", "smollm2:135m")
req_body = {
"model": model,
@@ -86,7 +95,6 @@ def _ollama_chat(user_text: str) -> str:
with urlopen(req, timeout=60) as resp:
data = json.loads(resp.read().decode("utf-8"))
except Exception as e:
# Let caller decide how to respond; include a useful hint in logs.
_log(f"Ollama request failed url={ollama_url!r} err={e!r}")
raise
@@ -97,6 +105,95 @@ def _ollama_chat(user_text: str) -> str:
return content.strip()
import re
# Sentence boundary: split on .!? followed by space or end, but not on
# abbreviations like "Dr." or "Mr." or decimals like "3.5".
_SENTENCE_END_RE = re.compile(
r'(?<=[.!?])\s+(?=[A-Z"\'])|(?<=[.!?])$'
)
def _split_sentences(text: str) -> list[str]:
"""Split text into sentences. Returns list of sentence strings."""
parts = _SENTENCE_END_RE.split(text.strip())
return [p.strip() for p in parts if p.strip()]
def _ollama_chat_streaming(user_text: str, on_sentence):
"""Stream Ollama response token-by-token, calling on_sentence(sentence_str)
each time a complete sentence is detected. Returns the full reply string.
This is the key to sub-5-second perceived latency: the robot starts speaking
the first sentence while the LLM is still generating the rest.
"""
ollama_url = os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434/api/chat")
model = os.environ.get("OLLAMA_MODEL", "smollm2:135m")
req_body = {
"model": model,
"stream": True,
"messages": [
{"role": "system", "content": "You are Jibo, a friendly home robot. Keep replies short and spoken."},
{"role": "user", "content": user_text},
],
}
req = Request(
ollama_url,
data=json.dumps(req_body).encode("utf-8"),
headers={"Content-Type": "application/json"},
method="POST",
)
full_reply = ""
buffer = ""
sentences_sent = 0
t0 = time.monotonic()
with urlopen(req, timeout=120) as resp:
# Ollama streams one JSON object per line
for raw_line in resp:
line = raw_line.decode("utf-8", errors="replace").strip()
if not line:
continue
try:
chunk = json.loads(line)
except json.JSONDecodeError:
continue
# Each chunk: {"message": {"content": "token"}, "done": false}
msg = chunk.get("message") or {}
token = msg.get("content") or ""
if token:
buffer += token
full_reply += token
# Check if buffer contains a complete sentence to flush
# Look for sentence-ending punctuation followed by a space or more tokens
sentences = _split_sentences(buffer)
if len(sentences) > 1:
# All but the last are complete sentences; last is still building
for s in sentences[:-1]:
elapsed = time.monotonic() - t0
_log(f" sentence #{sentences_sent} at {elapsed:.2f}s: {s[:120]!r}")
on_sentence(s)
sentences_sent += 1
buffer = sentences[-1]
if chunk.get("done"):
break
# Flush any remaining text as the final sentence
leftover = buffer.strip()
if leftover:
elapsed = time.monotonic() - t0
_log(f" sentence #{sentences_sent} (final) at {elapsed:.2f}s: {leftover[:120]!r}")
on_sentence(leftover)
return full_reply.strip()
def _short_err(e: BaseException) -> str:
s = str(e) or e.__class__.__name__
s = " ".join(s.split())
@@ -265,15 +362,55 @@ class Handler(BaseHTTPRequestHandler):
return
_log(f"{client} /text prompt_chars={len(text)} prompt={text[:200]!r}")
try:
reply = _ollama_chat(text)
_log(f"{client} /text ok reply_chars={len(reply)}")
_json_response(self, 200, {"reply": reply})
except URLError as e:
# Stream response: chunked transfer encoding with NDJSON
# Each line is a JSON object the robot can parse immediately.
# This replicates the original Jibo hub pattern where the robot
# starts acting on partial results while the server is still working.
self.send_response(200)
self.send_header("Content-Type", "application/x-ndjson")
self.send_header("Transfer-Encoding", "chunked")
self.end_headers()
t0 = time.monotonic()
def _send_chunk(obj: dict):
line = json.dumps(obj) + "\n"
data = line.encode("utf-8")
# HTTP chunked encoding: hex size + CRLF + data + CRLF
self.wfile.write(f"{len(data):x}\r\n".encode())
self.wfile.write(data)
self.wfile.write(b"\r\n")
self.wfile.flush()
def _on_sentence(sentence: str):
_send_chunk({"sentence": sentence, "done": False})
reply = _ollama_chat_streaming(text, _on_sentence)
elapsed = time.monotonic() - t0
_log(f"{client} /text ok reply_chars={len(reply)} elapsed={elapsed:.2f}s")
# Final chunk with full reply (for logging / legacy compat)
_send_chunk({"sentence": "", "done": True, "reply": reply})
# Terminate chunked encoding
self.wfile.write(b"0\r\n\r\n")
self.wfile.flush()
except (URLError, ConnectionRefusedError) as e:
_log(f"{client} /text ollama_unreachable err={_short_err(e)!r}")
_json_response(self, 200, {"reply": _ollama_down_reply(), "ollama_ok": False, "ollama_error": _short_err(e)})
except ConnectionRefusedError as e:
_log(f"{client} /text ollama_refused err={_short_err(e)!r}")
_json_response(self, 200, {"reply": _ollama_down_reply(), "ollama_ok": False, "ollama_error": _short_err(e)})
# Fallback: try to send a non-streaming error response.
# If headers already sent, write it as a chunk.
try:
err_reply = _ollama_down_reply()
err_obj = {"sentence": err_reply, "done": True, "reply": err_reply,
"ollama_ok": False, "ollama_error": _short_err(e)}
line = json.dumps(err_obj) + "\n"
data = line.encode("utf-8")
self.wfile.write(f"{len(data):x}\r\n".encode())
self.wfile.write(data)
self.wfile.write(b"\r\n0\r\n\r\n")
self.wfile.flush()
except Exception:
pass # headers may not have been sent yet
return
if self.path == "/v1/chat/audio":
@@ -371,7 +508,7 @@ def main():
"Ollama: "
+ os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434/api/chat")
+ " model="
+ os.environ.get("OLLAMA_MODEL", "phi3.5")
+ os.environ.get("OLLAMA_MODEL", "smollm2:135m")
)
_log("Ollama health check: curl -s http://127.0.0.1:11434/api/tags | head")
if not _whisper.available():