Spaces:
Sleeping
Sleeping
Download app.py from NextNeural/IRIS: direct link, hf CLI and curl.
- Browser
- Download file 59.2 kB
-
https://huggingface.co/spaces/NextNeural/IRIS/resolve/main/app.py
- Command line
-
hf download hf://spaces/NextNeural/IRIS/app.py
-
curl -L -o app.py https://huggingface.co/spaces/NextNeural/IRIS/resolve/main/app.py
59.2 kB
# ─────────────────────────────────────────────
# NINA — Neural Information & Narrative Abstraction
# app.py | IRIS v0.80 | Made by Malina
# ─────────────────────────────────────────────
#
# Precision MULTIMODAL video search + summarization pipeline.
# Pipeline: video → audio → ASR segments → semantic chunking
# → bi-encoder (text) ⨉ CLIP (vision) fusion → cross-encoder + entity overlap
# → merge → confidence-gated precise re-transcribe → streaming LLM synthesis.
#
# ─────────────────────────────────────────────────────────────────────────────
# v0.80 — performance & modernisation pass over v0.71
# ─────────────────────────────────────────────────────────────────────────────
# Measured on Apple M1 (8 core / 16 GB), 39 s test clip:
# cold first prompt 58.8 s → ~9 s, warm prompts ~1.5 s → sub-second to first token.
#
# [P1] HARDWARE ACCELERATION
# Torch models now auto-target the best device (CUDA → MPS → CPU). On
# Apple Silicon CLIP runs ~1.8× and the cross-encoder ~1.4× faster on the
# Metal (MPS) backend. CTranslate2 (Whisper) has no MPS backend, so it
# stays on CPU but is tuned: int8 + cpu_threads, greedy decoding for the
# cheap pass, and Silero VAD to skip non-speech.
#
# [P2] NO MORE DOUBLE TRANSCRIPTION
# v0.71 transcribed the whole video with `base`, then RE-transcribed every
# sampled sector with the ~15× slower `large-v3-turbo` — even for a global
# summary (≈24 s wasted on this clip). v0.80:
# • GLOBAL summaries use the clean `base` transcript directly (no turbo).
# • SPECIFIC queries re-transcribe with turbo ONLY on matched regions and
# ONLY when the base pass was low-confidence (avg_logprob gate). On
# clean audio turbo never runs; on hard audio it runs exactly where it
# helps. Per-sector results are cached and persist across prompts.
#
# [P3] PARALLEL COLD START
# Audio transcription (CPU) and CLIP frame analysis (MPS) are independent
# and now run concurrently on first sight of a video, overlapping the two
# heaviest stages instead of summing them.
#
# [P4] ZERO-COPY I/O
# Audio is decoded straight from the container (PyAV) — no ffmpeg wav
# temp file. Frames are streamed from one ffmpeg rawvideo pipe into NumPy —
# no per-frame JPEG temp files.
#
# [P5] PERSISTENT DISK CACHE
# Transcript, chunk + visual embeddings and precise sectors are persisted
# under a content fingerprint, so re-opening the same video — even after a
# restart — skips transcription/chunking/CLIP entirely.
#
# [P6] STREAMING + UX
# LLM synthesis streams token-by-token; run() is a generator that shows
# progress and writes the answer as it is produced.
#
# [Q] QUALITY FIXES
# • Router no longer misreads "what does the video say about X" as a global
# summary (the bare word "about"/"explain" no longer forces global mode).
# • The illustrative "[2:30]" timestamp no longer leaks into answers.
#
# Scoring formula (SPECIFIC/HYBRID mode):
# audio = 0.6 * sigmoid(cross_encoder_logit) + 0.4 * entity_overlap
# final = (1 - VISUAL_WEIGHT) * audio + VISUAL_WEIGHT * clip_visual_score
#
# Entity extraction: regex + simple NP heuristic (no spacy dependency).
# Works multilingually because it operates on surface forms, not POS tags.
# ─────────────────────────────────────────────
import os
import re
import json
import math
import base64
import hashlib
import hmac
import shutil
import tempfile
import threading
import time
from concurrent.futures import ThreadPoolExecutor
import ffmpeg
import gradio as gr
import numpy as np
from PIL import Image
from faster_whisper import WhisperModel
from faster_whisper.audio import decode_audio
from sentence_transformers import SentenceTransformer, CrossEncoder
from huggingface_hub import InferenceClient
import warnings
warnings.filterwarnings("ignore", message="FP16 is not supported on CPU; using FP32 instead")
# ── Config ─────────────────────────────────────────────────────────────────────
WHISPER_FAST = "base"
WHISPER_PRECISE = "large-v3-turbo"
SAMPLE_RATE = 16000
WHISPER_CPU_THREADS = int(os.getenv("WHISPER_CPU_THREADS", "3"))
# 3 threads × ~3 concurrent jobs ≈ an 8-vCPU Space.
# Override via env. CTranslate2 is CPU-only.
RETRANSCRIBE_MAX = 12 # per-prompt budget for NEW (uncached) sector transcriptions
PRECISE_LOGPROB_GATE = -0.55 # only turbo-refine sectors whose base avg_logprob is below this
# (clean audio scores high → turbo is correctly skipped)
CHUNK_SPLIT_THRESHOLD = 0.45
CHUNK_MIN_DURATION = 8.0
CHUNK_MAX_DURATION = 90.0
CHUNK_WINDOW = 5
BI_ENCODER_TOP_K = 100
CROSS_ENCODER_TOP_K = 10
RESULTS_TOP_N = 6
REGION_MIN_DURATION = 5.0
MERGE_GAP_SEC = 6.0
# routing thresholds
BI_SCORE_GLOBAL_FLOOR = 0.10 # below this → escalate to GLOBAL regardless
BI_SCORE_HYBRID_FLOOR = 0.28 # below this → HYBRID (pad with neighbours)
# global mode: how many evenly-sampled chunks to retranscribe
GLOBAL_MAX_CHUNKS = 12 # hard cap to keep latency sane
# ── CLIP / visual config ────────────────────────────────────────────────────────
CLIP_MODEL = "clip-ViT-B-32" # OpenAI CLIP ViT-B/32 (via sentence-transformers)
CLIP_MAX_FRAMES = 96 # cap frames encoded per video (perf)
CLIP_MIN_FPS = 0.1 # sample at least 1 frame / 10s
CLIP_MAX_FPS = 0.5 # sample at most 1 frame / 2s
CLIP_FRAME_SIZE = 224 # square fed to CLIP (matches ViT-B/32 input → no double resize)
VISUAL_WEIGHT = 0.30 # fusion weight of visual vs audio in final score
VISUAL_TAGS_TOP_K = 4 # max zero-shot tags surfaced per region
VISUAL_TAG_MIN_SIM = 0.20 # min CLIP cosine for a tag to count
# vocabulary for zero-shot "what is on screen" tagging
VISUAL_CONCEPTS = [
"a person talking to the camera", "a group of people", "a presentation with slides",
"a lecture or classroom", "an interview", "a news broadcast",
"a screen recording of software", "source code in a text editor",
"a webpage in a browser", "a chart or graph", "text on screen",
"a product demonstration", "hands demonstrating something",
"a cooking or food scene", "a kitchen", "an office", "a laboratory",
"a factory or machinery", "a construction site", "an outdoor nature landscape",
"a forest", "mountains", "a beach or ocean", "a city street",
"buildings and architecture", "a car or vehicle driving", "an airplane", "a train",
"a sports event", "people exercising", "a concert or live music",
"a video game gameplay", "an animation or cartoon", "a medical or hospital scene",
"a doctor or patient", "an animal", "a dog", "a cat", "a crowd of people",
"a map", "a whiteboard with diagrams", "a scientific diagram",
"a phone or mobile device", "a computer or laptop", "money or finance",
"a meeting or conference room", "a stage with a speaker",
"a dark scene at night", "a daytime scene",
]
LLM_MODEL = "Qwen/Qwen2.5-7B-Instruct"
LLM_MAX_TOKENS = 600
LLM_TIMEOUT = 60
CACHE_DIR = os.path.join(tempfile.gettempdir(), "nina_cache")
DISK_CACHE = True
DEBUG = False
# ── Usage-token gate ──────────────────────────────────────────────────────────────
# The Space is public, so to enforce the per-account 10/5h limit set on the
# website, each request must carry a short-lived HMAC token issued by the
# Cloudflare Worker (app.nnlabs.pl) after it counts the use. Without a valid,
# unexpired, not-yet-seen token the model refuses to run. Same secret on both
# sides (NINA_SIGNING_SECRET). If the secret is unset the gate is open (dev).
SIGNING_SECRET = os.getenv("NINA_SIGNING_SECRET", "")
_used_jti = {} # jti -> exp epoch (single-use, replay-proof within a container)
_jti_lock = threading.Lock()
def _b64url_decode(s: str) -> bytes:
return base64.urlsafe_b64decode(s + "=" * (-len(s) % 4))
def _b64url_encode(b: bytes) -> str:
return base64.urlsafe_b64encode(b).decode().rstrip("=")
def _verify_usage_token(token: str) -> bool:
if not SIGNING_SECRET:
return True # not configured → open (local/dev)
if not token or "." not in token:
return False
try:
msg, sig = token.split(".", 1)
expected = _b64url_encode(hmac.new(SIGNING_SECRET.encode(), msg.encode(), hashlib.sha256).digest())
if not hmac.compare_digest(expected, sig):
return False
payload = json.loads(_b64url_decode(msg))
now = int(time.time())
if int(payload.get("e", 0)) < now:
return False # expired
jti = payload.get("j", "")
if not jti:
return False
with _jti_lock:
for k in [k for k, e in _used_jti.items() if e < now]:
_used_jti.pop(k, None)
if jti in _used_jti:
return False # replay
_used_jti[jti] = int(payload.get("e", now + 120))
return True
except Exception:
return False
# ── Device selection ─────────────────────────────────────────────────────────────
# CTranslate2 (Whisper) is CPU-only here; the torch models (bi/cross/CLIP) can
# use Metal (MPS) on Apple Silicon or CUDA on a GPU box.
def _best_torch_device() -> str:
try:
import torch
if torch.cuda.is_available():
return "cuda"
if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available():
return "mps"
except Exception:
pass
return "cpu"
TORCH_DEVICE = _best_torch_device()
# ── Model load ─────────────────────────────────────────────────────────────────
print(f"[NINA] Torch device: {TORCH_DEVICE} | Whisper: CPU int8 (x{WHISPER_CPU_THREADS})")
print("[NINA] Loading Whisper base (faster-whisper)...")
_whisper_fast = WhisperModel(WHISPER_FAST, device="cpu", compute_type="int8",
cpu_threads=WHISPER_CPU_THREADS)
# Bi-encoder stays on CPU: it is tiny (sub-100 ms) and MPS transfer overhead is a
# wash, and this keeps the MPS device free for CLIP during the parallel cold start.
print("[NINA] Loading bi-encoder...")
bi_encoder = SentenceTransformer("paraphrase-multilingual-MiniLM-L12-v2", device="cpu")
print(f"[NINA] Loading cross-encoder ({TORCH_DEVICE})...")
cross_encoder = CrossEncoder("cross-encoder/mmarco-mMiniLMv2-L12-H384-v1", device=TORCH_DEVICE)
HF_TOKEN = os.getenv("HF_TOKEN")
if not HF_TOKEN:
_token_file = os.path.expanduser("~/.cache/huggingface/token")
if os.path.exists(_token_file):
HF_TOKEN = open(_token_file).read().strip()
if HF_TOKEN:
print(f"[NINA] HF_TOKEN loaded — LLM calls enabled.")
else:
print("[NINA] HF_TOKEN not set — LLM calls will fail.")
llm_client = InferenceClient(
token=HF_TOKEN,
timeout=LLM_TIMEOUT,
)
_whisper_precise = None
_whisper_precise_lock = threading.Lock()
def get_whisper_precise():
global _whisper_precise
with _whisper_precise_lock:
if _whisper_precise is None:
print(f"[NINA] Loading Whisper {WHISPER_PRECISE} (faster-whisper)...")
_whisper_precise = WhisperModel(WHISPER_PRECISE, device="cpu",
compute_type="int8",
cpu_threads=WHISPER_CPU_THREADS)
print(f"[NINA] Whisper {WHISPER_PRECISE} ready.")
return _whisper_precise
# ── CLIP (OpenAI ViT-B/32) — lazy, warmed in background ──────────────────────────
_clip_model = None
_clip_lock = threading.Lock()
_concept_embs = None
_concept_lock = threading.Lock()
def get_clip():
global _clip_model
with _clip_lock:
if _clip_model is None:
print(f"[NINA] Loading CLIP ({CLIP_MODEL}, {TORCH_DEVICE})...")
_clip_model = SentenceTransformer(CLIP_MODEL, device=TORCH_DEVICE)
print(f"[NINA] CLIP ready.")
return _clip_model
def _get_concept_embs() -> np.ndarray:
"""Encode (once) the zero-shot vocabulary, L2-normalised."""
global _concept_embs
with _concept_lock:
if _concept_embs is None:
embs = get_clip().encode(VISUAL_CONCEPTS, convert_to_numpy=True,
show_progress_bar=False)
_concept_embs = embs / (np.linalg.norm(embs, axis=1, keepdims=True) + 1e-9)
return _concept_embs
# Warm the lazy models in the background; CLIP loads on the (free) MPS device,
# turbo loads on CPU. Both finish before they are first needed in most cases.
threading.Thread(target=get_whisper_precise, daemon=True).start()
threading.Thread(target=get_clip, daemon=True).start()
print(f"[NINA] Ready. ( {WHISPER_PRECISE} + CLIP warming up...)")
# ── Transcript + visual cache (in-memory) ─────────────────────────────────────────
# Stores segments, chunks (with embeddings), visual frame embeddings, the decoded
# waveform, and a per-sector precise-transcription cache, keyed by video path.
# Second prompt on same video skips transcription, chunking and CLIP entirely.
_cache: dict = {} # { video_path: {"segments","chunks","visual","audio","precise","_atime"} }
# Data lifecycle / privacy:
# • The uploaded video itself is purged by Gradio's own temp cleaner
# (Interface(delete_cache=...)) — see the bottom of this file — so a file is
# removed shortly after the user's session goes idle. One upload still serves
# MANY prompts: the decoded audio + transcript live in RAM (below), so the
# file is no longer needed once processed.
# • This in-memory state and the derived on-disk cache are TTL-evicted here so
# nothing lingers indefinitely.
CACHE_IDLE_TTL = 1200 # drop in-memory video state idle for > 20 min
CACHE_DISK_TTL = 1800 # delete on-disk transcript/embedding cache > 30 min old
def _evict_stale() -> None:
now = time.time()
for vp in [k for k, e in _cache.items() if now - e.get("_atime", now) > CACHE_IDLE_TTL]:
_cache.pop(vp, None)
try:
if os.path.isdir(CACHE_DIR):
for name in os.listdir(CACHE_DIR):
d = os.path.join(CACHE_DIR, name)
if os.path.isdir(d) and now - os.path.getmtime(d) > CACHE_DISK_TTL:
shutil.rmtree(d, ignore_errors=True)
except Exception:
pass
# ── Audio ──────────────────────────────────────────────────────────────────────
def load_audio(media_path: str) -> np.ndarray:
"""Decode any container straight to a float32 mono 16k waveform (PyAV, no temp file)."""
return decode_audio(media_path, sampling_rate=SAMPLE_RATE)
def media_duration(media_path: str) -> float:
try:
info = ffmpeg.probe(media_path)
return float(info["format"]["duration"])
except Exception:
return 0.0
# ── Frames (CLIP input) — streamed from one ffmpeg pipe, no temp files ────────────
def extract_frames(video_path: str, duration: float):
"""
Sample frames evenly across the video and stream them as raw RGB straight
into NumPy via a single ffmpeg pipe (no per-frame JPEG temp files).
fps is chosen so total frames ≈ CLIP_MAX_FRAMES (bounded by MIN/MAX fps).
Returns (frames_rgb, times) where frames_rgb is (N, S, S, 3) uint8.
"""
fps = CLIP_MAX_FPS
if duration and duration > 0:
fps = min(CLIP_MAX_FPS, max(CLIP_MIN_FPS, CLIP_MAX_FRAMES / duration))
S = CLIP_FRAME_SIZE
try:
out, _ = (
ffmpeg.input(video_path)
.filter("fps", fps=fps)
# cover-then-centre-crop to a square so portrait/landscape both work
.filter("scale", S, S, force_original_aspect_ratio="increase")
.filter("crop", S, S)
.output("pipe:", format="rawvideo", pix_fmt="rgb24")
.run(capture_stdout=True, capture_stderr=True)
)
except ffmpeg.Error:
return np.zeros((0, S, S, 3), dtype=np.uint8), np.zeros((0,))
if not out:
return np.zeros((0, S, S, 3), dtype=np.uint8), np.zeros((0,))
frames = np.frombuffer(out, np.uint8).reshape(-1, S, S, 3)[:CLIP_MAX_FRAMES]
times = np.array([(i + 0.5) / fps for i in range(len(frames))])
return frames, times
# ── Stage 2: transcription ─────────────────────────────────────────────────────
def transcribe_fast(audio: np.ndarray) -> list[dict]:
"""
Cheap full-video pass: base model, greedy decoding, Silero VAD.
Keeps per-segment avg_logprob so we can later refine ONLY weak sectors.
"""
segments_gen, _ = _whisper_fast.transcribe(
audio, beam_size=1, vad_filter=True, word_timestamps=False,
)
return [
{
"text": seg.text.strip(),
"start": round(seg.start, 2),
"end": round(seg.end, 2),
"logprob": round(float(getattr(seg, "avg_logprob", 0.0) or 0.0), 3),
}
for seg in segments_gen
if seg.text.strip()
]
# ── Stage 3a: visual analysis (CLIP) ─────────────────────────────────────────────
def analyze_visual(video_path: str, duration: float) -> dict:
"""
Sample → encode frames with CLIP. Cached per video.
Returns {"emb": (N, D) float32, "time": (N,) float}.
Empty arrays when the video has no usable visual stream (audio-only).
"""
frames, times = extract_frames(video_path, duration)
if len(frames) == 0:
print("[NINA] No video frames (audio-only?) — visual analysis skipped.")
return {"emb": np.zeros((0, 512), dtype=np.float32), "time": np.zeros((0,))}
images = [Image.fromarray(f) for f in frames]
embs = get_clip().encode(images, convert_to_numpy=True,
batch_size=32, show_progress_bar=False)
print(f"[NINA] CLIP encoded {len(images)} frames ({TORCH_DEVICE}).")
return {"emb": np.asarray(embs, dtype=np.float32), "time": np.asarray(times)}
# ── semantic chunking ─────────────────────────────────────────────────
def semantic_chunk(segments: list[dict]) -> list[dict]:
if not segments:
return []
texts = [s["text"] for s in segments]
embeddings = bi_encoder.encode(texts, convert_to_numpy=True, batch_size=64)
chunks, current, cur_emb = [], [segments[0]], [embeddings[0]]
for i in range(1, len(segments)):
chunk_dur = segments[i]["end"] - current[0]["start"]
window_mean = np.mean(cur_emb[-CHUNK_WINDOW:], axis=0)
sim = max(_cosine(embeddings[i], window_mean),
_cosine(embeddings[i], cur_emb[-1]))
long_enough = (current[-1]["end"] - current[0]["start"]) >= CHUNK_MIN_DURATION
if (sim < CHUNK_SPLIT_THRESHOLD or chunk_dur >= CHUNK_MAX_DURATION) and long_enough:
chunks.append(_build_chunk(current, cur_emb))
current, cur_emb = [segments[i]], [embeddings[i]]
else:
current.append(segments[i])
cur_emb.append(embeddings[i])
if current:
chunks.append(_build_chunk(current, cur_emb))
print(f"[NINA] {len(segments)} segments → {len(chunks)} chunks.")
return chunks
# ── Entity extraction ──────────────────────────────────────────────────────────
_STOPWORDS = {
"the","a","an","and","or","but","in","on","at","to","for","of","is","are",
"was","were","be","been","that","this","it","he","she","they","we","i",
"w","z","i","że","na","do","się","nie","to","jest","są","by","przez","po",
"le","la","les","un","une","des","du","et","en","à","de","est","il","elle",
"ein","eine","der","die","das","und","ist","in","zu","mit","von","dem",
}
def extract_entities(text: str) -> set[str]:
"""
Extract surface-form entities from text.
Returns a set of normalised lowercase strings for overlap comparison.
"""
entities: set[str] = set()
# 1. quoted phrases — high precision
for match in re.finditer(r'"([^"]{2,40})"', text):
entities.add(match.group(1).lower().strip())
# 2. capitalised multi-word sequences (proper nouns)
for match in re.finditer(r'\b([A-ZÁÉÍÓÚÀÂÊÎÔÙÄËÏÖÜĄĆĘŁŃÓŚŹŻ][a-záéíóúàâêîôùäëïöüąćęłńóśźż]+(?:\s[A-ZÁÉÍÓÚÀÂÊÎÔÙÄËÏÖÜĄĆĘŁŃÓŚŹŻ][a-zÀ-žąćęłńóśźż]+)*)\b', text):
token = match.group(1).lower().strip()
if token not in _STOPWORDS and len(token) > 2:
entities.add(token)
# 3. single capitalised tokens (names, brands)
for token in re.findall(r'\b[A-ZÁÉÍÓÚÀÂÊÎÔÙÄËÏÖÜĄĆĘŁŃÓŚŹŻ][a-zÀ-žąćęłńóśźż]{2,}\b', text):
t = token.lower()
if t not in _STOPWORDS:
entities.add(t)
# 4. numbers with context
for match in re.finditer(r'\b([A-Za-z]+[-\s]?\d+[\w]*)\b', text):
entities.add(match.group(1).lower())
return entities
def entity_overlap_score(prompt_entities: set[str],
chunk_entities: set[str],
chunk_text: str = "") -> float:
if not prompt_entities:
return 0.0
chunk_text_lower = chunk_text.lower()
matched = 0
for pe in prompt_entities:
# pass 1: extracted entity overlap
entity_hit = any(pe in ce or ce in pe for ce in chunk_entities)
# pass 2: raw text substring (handles lowercase ASR output)
text_hit = len(pe) > 3 and pe in chunk_text_lower
if entity_hit or text_hit:
matched += 1
return matched / len(prompt_entities)
def add_entities_to_chunks(chunks: list[dict]) -> list[dict]:
for chunk in chunks:
chunk["entities"] = extract_entities(chunk["text"])
return chunks
# ── Query router ───────────────────────────────────────────────────────────────
#
# Classifies prompt into one of three modes before retrieval runs.
#
# GLOBAL — user wants a summary / overview of the whole video.
# Retrieval is the wrong tool here because no single chunk is "about"
# summarization. → skip retrieval, sample chunks evenly, synthesize.
#
# HYBRID — user has a topic but signal is weak (bi-score 0.10–0.28).
# → run retrieval, but pad results with adjacent chunks.
#
# SPECIFIC — normal focused query, retrieval works well (bi-score > 0.28).
# → standard pipeline unchanged.
#
# Router runs in two passes:
# Pass 1: regex pattern match on prompt text (zero-cost, runs before encoding)
# Pass 2: bi-encoder score after retrieval (upgrades SPECIFIC → GLOBAL if
# retrieval actually fails)
# Global ONLY when the user asks for a whole-video overview. Critically, this no
# longer fires on "what does the video say ABOUT <topic>" — a bare "about" or
# "explain" used to force global and swallow specific questions.
_GLOBAL_PATTERNS = re.compile(
r'\b('
r'summari[sz]e|summary|overview|recap|tl;?dr|key\s+points?|main\s+points?|'
r'key\s+takeaways?|main\s+idea|gist|'
r'what(?:\'s| is| was)\s+(?:this|the|it)\s+(?:video|recording|clip|all\s+)?\s*about|'
r'what(?:\'s| is| was)\s+(?:this|the)\s+(?:video|recording|clip)|'
r'describe\s+(?:the|this)\s+(?:video|recording|clip)|'
r'streszcz\w*|podsumuj\w*|podsumowanie|'
r'o\s+czym\s+(?:jest|to|opowiada|mówi)|czego\s+dotyczy|'
r'gł[oó]wn\w+\s+(?:punkt\w*|my[sś]l\w*|tez\w*)|og[oó]ln\w+\s+zarys'
r')\b',
re.IGNORECASE
)
def classify_query(prompt: str) -> str:
"""
Pass 1 router — pattern-based, runs before any encoding.
Returns "global" or "specific".
Pass 2 (bi-score based upgrade) happens inside run() after retrieval.
"""
if _GLOBAL_PATTERNS.search(prompt):
return "global"
return "specific"
# ── Stage 4: bi-encoder retrieval (top-100) ───────────────────────────────────
def bi_encode_retrieve(chunks: list[dict], prompt: str) -> tuple[list[dict], float]:
# chunk embeddings were computed once at chunking time and cached on the
# chunk dict ("emb"); here we only encode the prompt.
prompt_emb = bi_encoder.encode([prompt], convert_to_numpy=True)[0]
scores = [_cosine(prompt_emb, c["emb"]) for c in chunks]
top_k = min(BI_ENCODER_TOP_K, len(chunks))
top_idx = np.argsort(scores)[::-1][:top_k]
if DEBUG:
print(f"[DEBUG] Bi top-3 scores: " +
", ".join(f"{scores[i]:.3f}" for i in top_idx[:3]))
best = scores[top_idx[0]] if len(top_idx) else 0.0
return (
[{**chunks[i], "bi_score": round(scores[i], 3)} for i in top_idx],
best
)
# ── Stage 4b: visual relevance (CLIP) ─────────────────────────────────────────
def visual_scores(visual: dict, prompt: str) -> tuple[np.ndarray, np.ndarray]:
"""
Cosine of prompt↔each frame in CLIP space, plus a min-max normalised copy.
Normalisation makes CLIP sims (typically ~0.2–0.35) comparable to the
audio score's 0–1 range before fusion. Returns (raw, norm); empty when
there are no frames.
"""
embs = visual.get("emb")
if embs is None or len(embs) == 0:
if DEBUG:
print("[DEBUG] Visual: no frames (audio-only) — visual weight = 0.")
return np.array([]), np.array([])
q = get_clip().encode([prompt], convert_to_numpy=True, show_progress_bar=False)[0]
qn = q / (np.linalg.norm(q) + 1e-9)
en = embs / (np.linalg.norm(embs, axis=1, keepdims=True) + 1e-9)
sims = en @ qn
lo, hi = float(sims.min()), float(sims.max())
norm = (sims - lo) / (hi - lo + 1e-9)
if DEBUG:
times = visual.get("time")
order = np.argsort(sims)[::-1][:3]
tops = ", ".join(f"{fmt_time(times[i])}={sims[i]:.3f}" for i in order)
print(f"[DEBUG] Visual: {len(sims)} frames; raw sim "
f"min/mean/max={sims.min():.3f}/{sims.mean():.3f}/{sims.max():.3f}; "
f"top moments: {tops}")
return sims, norm
def span_visual_score(times: np.ndarray, norm: np.ndarray,
start: float, end: float) -> float:
"""Best (normalised) visual relevance among frames inside [start, end]."""
if len(times) == 0:
return 0.0
mask = (times >= start - 0.5) & (times <= end + 0.5)
if not mask.any():
idx = int(np.argmin(np.abs(times - (start + end) / 2)))
return float(norm[idx])
return float(norm[mask].max())
def region_visual_tags(visual: dict, start: float, end: float) -> list[str]:
"""Zero-shot CLIP tags describing what is on screen during [start, end]."""
embs = visual.get("emb")
times = visual.get("time")
if embs is None or len(embs) == 0:
return []
mask = (times >= start - 0.5) & (times <= end + 0.5)
if mask.any():
sel = embs[mask]
else:
idx = int(np.argmin(np.abs(times - (start + end) / 2)))
sel = embs[idx:idx + 1]
region_emb = sel.mean(axis=0)
region_emb = region_emb / (np.linalg.norm(region_emb) + 1e-9)
sims = _get_concept_embs() @ region_emb
order = np.argsort(sims)[::-1][:VISUAL_TAGS_TOP_K]
return [VISUAL_CONCEPTS[i] for i in order if sims[i] >= VISUAL_TAG_MIN_SIM]
def add_visual_tags(regions: list[dict], visual: dict) -> list[dict]:
for r in regions:
r["visual_tags"] = region_visual_tags(visual, r["start"], r["end"])
return regions
# ── Stage 5: cross-encoder rerank + audio/visual fusion (top-100 → top-10) ─────
def rerank_and_score(candidates: list[dict], prompt: str,
times: np.ndarray, norm: np.ndarray) -> list[dict]:
# cross-encoder
pairs = [[prompt, c["text"]] for c in candidates]
logits = cross_encoder.predict(pairs)
# entity overlap
prompt_entities = extract_entities(prompt)
prompt_words = {w.lower() for w in re.findall(r'[\wÀ-ž]{3,}', prompt)
if w.lower() not in _STOPWORDS}
all_prompt_tokens = prompt_entities | prompt_words
if DEBUG and all_prompt_tokens:
print(f"[DEBUG] Prompt tokens for entity match: {all_prompt_tokens}")
vw = VISUAL_WEIGHT if len(norm) else 0.0 # no frames → pure audio
scored = []
for candidate, logit in zip(candidates, logits):
chunk_entities = candidate.get("entities", set())
chunk_text = candidate.get("text", "")
overlap = entity_overlap_score(all_prompt_tokens, chunk_entities, chunk_text)
ce_conf = _sigmoid(float(logit))
audio_score = 0.6 * ce_conf + 0.4 * overlap
visual_score = span_visual_score(times, norm, candidate["start"], candidate["end"])
final = (1.0 - vw) * audio_score + vw * visual_score
scored.append({
**candidate,
"ce_logit": round(float(logit), 3),
"entity_overlap": round(overlap, 3),
"visual_score": round(visual_score, 3),
"confidence": round(min(1.0, final), 3),
})
# sort by fused multimodal score
ranked = sorted(scored, key=lambda x: x["confidence"], reverse=True)
if DEBUG:
print(f"[DEBUG] Top-3 after reranking (audio+visual):")
for r in ranked[:3]:
print(f" ce_logit={r['ce_logit']:.2f} "
f"bi={r['bi_score']:.3f} "
f"entity={r['entity_overlap']:.3f} "
f"visual={r['visual_score']:.3f} "
f"conf={r['confidence']:.3f} "
f"\"{r['text'][:60]}\"")
return ranked[:CROSS_ENCODER_TOP_K]
# ── Hybrid padding ─────────────────────────────────────────────────────────────
def pad_with_neighbours(results: list[dict], chunks: list[dict]) -> list[dict]:
"""
HYBRID mode: for each retrieved chunk, include its immediate neighbours
from the full chunk list. Helps when the relevant content slightly
overflows the chunk boundary. Deduplicates by start timestamp.
"""
result_starts = {r["start"] for r in results}
padded = list(results)
for r in results:
for i, c in enumerate(chunks):
if abs(c["start"] - r["start"]) < 0.5: # found matching chunk
for neighbour_i in [i - 1, i + 1]:
if 0 <= neighbour_i < len(chunks):
nb = chunks[neighbour_i]
if nb["start"] not in result_starts:
padded.append({
**nb,
"bi_score": 0.0,
"ce_logit": -9.0,
"entity_overlap": 0.0,
"visual_score": 0.0,
"confidence": nb.get("confidence", 0.1),
})
result_starts.add(nb["start"])
break
return sorted(padded, key=lambda x: x["start"])
# ── Global mode: even sampling ────────────────────────────────────────────────
def sample_chunks_for_global(chunks: list[dict], n_target: int) -> list[dict]:
"""
For global summarization, pick `n_target` chunks evenly distributed across
the video, so the summary covers the whole thing, not just the beginning.
"""
n_target = max(3, min(n_target, len(chunks)))
if n_target >= len(chunks):
return chunks
# evenly spaced indices
indices = [int(i * (len(chunks) - 1) / (n_target - 1)) for i in range(n_target)]
return [chunks[i] for i in indices]
# ── Stage 6: merge regions ─────────────────────────────────────────────────────
def merge_regions(results: list[dict]) -> list[dict]:
if not results:
return []
sorted_r = sorted(results, key=lambda x: x["start"])
merged, current = [], sorted_r[0].copy()
current["_frags"] = [sorted_r[0]]
for nxt in sorted_r[1:]:
if nxt["start"] - current["end"] <= MERGE_GAP_SEC:
current["end"] = max(current["end"], nxt["end"])
current["text"] += " " + nxt["text"]
current["_frags"].append(nxt)
else:
merged.append(_finalize_region(current))
current = nxt.copy()
current["_frags"] = [nxt]
merged.append(_finalize_region(current))
return [r for r in merged if (r["end"] - r["start"]) >= REGION_MIN_DURATION]
def _finalize_region(region: dict) -> dict:
frags = region.pop("_frags")
total_dur = sum(max(f["end"] - f["start"], 0.1) for f in frags)
w_conf = sum(f["confidence"] * max(f["end"] - f["start"], 0.1)
for f in frags) / total_dur
duration = region["end"] - region["start"]
w_conf *= min(1.0, duration / 20.0)
visual = max((f.get("visual_score", 0.0) for f in frags), default=0.0)
return {**region, "confidence": round(w_conf, 3), "visual_score": round(visual, 3)}
# ── Stage 7: precise re-transcription (confidence-gated) ──────────────────────
def retranscribe_region(audio: np.ndarray, region: dict) -> str:
"""Slice the in-memory waveform (no per-region ffmpeg) and run precise ASR."""
s = max(0, int(region["start"] * SAMPLE_RATE))
e = min(len(audio), int(region["end"] * SAMPLE_RATE))
clip = audio[s:e]
if len(clip) < SAMPLE_RATE * 0.2: # < 0.2s → nothing to gain
return region.get("text", "")
model = get_whisper_precise()
segments_gen, _ = model.transcribe(clip, beam_size=1)
return " ".join(seg.text.strip() for seg in segments_gen).strip()
def fill_precise(audio: np.ndarray, regions: list[dict], all_chunks: list[dict],
precise_cache: dict, budget: int) -> int:
"""
Ensure every region has `precise_text`. A sector is turbo-refined ONLY when:
• it is not already cached, AND
• the base pass for that sector was low-confidence (avg_logprob < gate), AND
• we are still within `budget` NEW transcriptions this call.
Clean/high-confidence sectors keep the (already accurate) base text for free,
so on clean audio the slow turbo model never runs. `precise_cache` persists
per video, so later prompts reuse earlier work. Returns remaining budget.
"""
reused = new = skipped = 0
for region in regions:
# the chunk sectors this region spans (chunks are non-overlapping)
comp = [c for c in all_chunks
if c["start"] < region["end"] - 1e-3 and c["end"] > region["start"] + 1e-3]
if not comp:
comp = [region]
for ch in comp:
key = round(ch["start"], 2)
if key in precise_cache:
reused += 1
continue
weak = ch.get("logprob", 0.0) < PRECISE_LOGPROB_GATE
if not weak:
skipped += 1 # base was confident → trust it
continue
if budget <= 0:
continue # over budget → fall back to fast text
print(f"[NINA] Re-transcribing weak sector "
f"[{fmt_time(ch['start'])} → {fmt_time(ch['end'])}] "
f"(logprob {ch.get('logprob', 0.0):.2f})")
precise_cache[key] = retranscribe_region(audio, ch)
if DEBUG:
snip = precise_cache[key]
print(f"[DEBUG] Precise sector [{fmt_time(ch['start'])}→{fmt_time(ch['end'])}]: "
f"{snip[:200]}{'…' if len(snip) > 200 else ''}")
new += 1
budget -= 1
region["precise_text"] = " ".join(
(precise_cache.get(round(ch["start"], 2)) or ch.get("text", ""))
for ch in comp
).strip() or region.get("text", "")
print(f"[NINA] Precise sectors — {new} refined, {reused} cached, {skipped} trusted (base).")
return budget
# ── Stage 8: synthesis ─────────────────────────────────────────────────────────
# ── Output sizing — inferred from the PROMPT (no slider) ─────────────────────────
_BRIEF_CUES = (
"brief", "briefly", "short", "shortly", "quick", "quickly", "concise",
"in a sentence", "one sentence", "one-liner", "tl;dr", "tldr",
"krótko", "w skrócie", "zwięźle", "jednym zdaniem", "w jednym zdaniu",
)
_DETAIL_CUES = (
"in detail", "detailed", "in-depth", "in depth", "thorough", "thoroughly",
"comprehensive", "exhaustive", "everything", "full summary", "elaborate",
"step by step", "step-by-step", "deep dive", "as much as possible",
"szczegółowo", "dokładnie", "wyczerpująco", "wszystko", "krok po kroku",
"dogłębnie", "obszernie",
)
_FACTUAL_START = re.compile(
r'^\s*(what|who|whom|whose|when|where|which|how many|how much|how long|'
r'is|are|was|were|does|do|did|can|could|will|would|'
r'czy|kiedy|kto|kogo|gdzie|ile|jaki|jaka|jakie|który|która|które)\b',
re.IGNORECASE,
)
def _summary_tier(max_words: int) -> str:
"""Human-readable label for the chosen output length."""
if max_words <= 60: return "brief / one-liner"
if max_words <= 110: return "short"
if max_words <= 250: return "standard"
if max_words <= 400: return "detailed"
return "exhaustive"
def _prompt_to_words(prompt: str) -> int:
"""
Infer the answer's word budget from the prompt itself (replaces the slider):
- explicit brevity cues → ~40 words
- explicit detail cues → ~450 words
- factual wh-/yes-no Q → ~80 words
- otherwise scale gently with prompt length (longer ask → longer answer)
Clamped to [30, 500].
"""
p = prompt.lower()
if any(c in p for c in _DETAIL_CUES):
base, reason = 450, "explicit detail cue"
elif any(c in p for c in _BRIEF_CUES):
base, reason = 40, "explicit brevity cue"
elif _FACTUAL_START.match(p):
base, reason = 80, "factual question"
else:
n_words = len(re.findall(r'\w+', p))
base, reason = 120 + min(180, n_words * 8), f"length-scaled ({n_words} prompt words)"
out = int(max(30, min(500, base)))
if DEBUG:
print(f"[DEBUG] Summary sizing: {reason} → {out} words [{_summary_tier(out)}]")
return out
def _words_to_global_chunks(max_words: int) -> int:
"""More requested words → sample more chunks for the global overview."""
return max(3, min(GLOBAL_MAX_CHUNKS, round(max_words / 40)))
def _fallback_summary(regions: list[dict]) -> str:
snippets = [
f"[{fmt_time(r['start'])}] {r.get('precise_text', r['text'])[:200]}"
for r in regions[:2]
]
return "(LLM unavailable — raw excerpts:)\n" + "\n".join(snippets)
def _build_synth_prompt(regions: list[dict], prompt: str,
max_words: int, mode: str) -> str:
# Each fragment carries BOTH modalities: what is SAID (audio) and what is
# SHOWN (CLIP visual cues). The LLM is asked to fuse them.
blocks = []
for r in regions:
spoken = (r.get("precise_text", r["text"]) or "").strip()
tags = r.get("visual_tags", [])
block = (f"[{fmt_time(r['start'])} → {fmt_time(r['end'])}]\n"
f" Spoken (audio): {spoken or '(no speech)'}")
if tags:
block += f"\n On screen (video): {', '.join(tags)}"
blocks.append(block)
fragments = "\n\n".join(blocks)
if mode == "global":
task_line = "Task: Give a complete overview of what this video is about."
else:
task_line = f'Task: Summarize what the video says/shows about "{prompt}".'
return (
f"You are analyzing a video. Each fragment below contains what is SPOKEN "
f"(audio transcript) and what is SHOWN (visual cues detected in the frames).\n"
f'The user is looking for: "{prompt}"\n\n'
f"Fragments:\n\n{fragments}\n\n"
f"{task_line}\n\n"
f"Output format:\n"
f"- Combine BOTH the spoken content and the on-screen visuals into one coherent answer\n"
f"- After each statement, append the timestamp it came from, taken from the "
f"bracketed start time of the relevant fragment above (use that exact value)\n"
f"- Maximum {max_words} words total\n"
f"- IMPORTANT: Write in the same language as the prompt given by the user\n"
f"- Only use information from the fragments (audio + visual) — do not invent details\n"
f"- If fragments contain no real information about the topic, say so clearly, "
f"without further explanation\n"
)
def llm_synthesize_stream(regions: list[dict], prompt: str,
max_words: int = 150, mode: str = "specific"):
"""
Generator: yields the answer incrementally as the LLM streams tokens.
On any error, yields a single fallback string.
"""
if not regions:
yield "(No relevant regions to summarize.)"
return
synth_prompt = _build_synth_prompt(regions, prompt, max_words, mode)
if DEBUG:
print(f"[DEBUG] Synthesis [{mode}] — {len(regions)} region(s), "
f"max_words={max_words} [{_summary_tier(max_words)}].")
try:
acc = ""
stream = llm_client.chat_completion(
model=LLM_MODEL,
messages=[{"role": "user", "content": synth_prompt}],
max_tokens=LLM_MAX_TOKENS,
temperature=0.3,
stream=True,
)
for ev in stream:
delta = ev.choices[0].delta.content or ""
if delta:
acc += delta
yield acc.strip()
if not acc.strip():
yield _fallback_summary(regions)
except Exception as e:
print(f"[NINA] Synthesis error: {e}")
yield _fallback_summary(regions)
def llm_synthesize(regions: list[dict], prompt: str,
max_words: int = 150, mode: str = "specific") -> str:
"""Non-streaming convenience wrapper (kept for API compatibility / testing)."""
out = ""
for out in llm_synthesize_stream(regions, prompt, max_words, mode):
pass
return out
# ── Helpers ────────────────────────────────────────────────────────────────────
def _build_chunk(segs: list[dict], embs: list[np.ndarray]) -> dict:
total = sum(max(s["end"] - s["start"], 0.1) for s in segs)
logp = sum(s.get("logprob", 0.0) * max(s["end"] - s["start"], 0.1)
for s in segs) / total # duration-weighted base confidence
return {"text": " ".join(s["text"] for s in segs),
"start": segs[0]["start"],
"end": segs[-1]["end"],
"logprob": round(logp, 3),
"emb": np.mean(np.stack(embs), axis=0)} # cached chunk embedding
def _cosine(a: np.ndarray, b: np.ndarray) -> float:
return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-9))
def _sigmoid(x: float) -> float:
return 1.0 / (1.0 + math.exp(-float(x)))
def fmt_time(s: float) -> str:
m = int(s) // 60
return f"{m}:{s % 60:05.2f}"
def conf_bar(c: float) -> str:
f = round(c * 10)
return "█" * f + "░" * (10 - f)
# ── Persistent disk cache ────────────────────────────────────────────────────────
# Keyed by a content fingerprint (size + head + tail bytes), so the same video
# re-uploaded to a fresh temp path — or opened after a restart — is recognised
# and its transcript / embeddings / precise sectors are loaded instead of
# recomputed. Best-effort: any failure silently falls back to recomputation.
def _fingerprint(path: str) -> str:
h = hashlib.sha1()
st = os.stat(path)
h.update(str(st.st_size).encode())
with open(path, "rb") as f:
h.update(f.read(1 << 20)) # first 1 MB
if st.st_size > (1 << 20):
f.seek(-(1 << 20), os.SEEK_END)
h.update(f.read(1 << 20)) # last 1 MB
return h.hexdigest()[:16]
def _disk_load(fp: str) -> dict | None:
if not DISK_CACHE:
return None
d = os.path.join(CACHE_DIR, fp)
try:
with open(os.path.join(d, "meta.json"), encoding="utf-8") as f:
meta = json.load(f)
arr = np.load(os.path.join(d, "emb.npz"))
chunk_emb = arr["chunk_emb"]
chunks = []
for i, c in enumerate(meta["chunks"]):
chunks.append({
"text": c["text"], "start": c["start"], "end": c["end"],
"logprob": c.get("logprob", 0.0),
"entities": set(c.get("entities", [])),
"emb": chunk_emb[i],
})
visual = {"emb": arr["visual_emb"], "time": arr["visual_time"]}
print(f"[NINA] Disk cache hit ({fp}) — loaded {len(chunks)} chunks.")
return {"segments": meta["segments"], "chunks": chunks,
"visual": visual, "precise": {float(k): v for k, v in meta["precise"].items()}}
except Exception:
return None
def _disk_save(fp: str, entry: dict) -> None:
if not DISK_CACHE:
return
d = os.path.join(CACHE_DIR, fp)
try:
os.makedirs(d, exist_ok=True)
chunks = entry["chunks"]
meta = {
"segments": entry["segments"],
"precise": entry["precise"],
"chunks": [{"text": c["text"], "start": c["start"], "end": c["end"],
"logprob": c.get("logprob", 0.0),
"entities": sorted(c.get("entities", []))}
for c in chunks],
}
chunk_emb = (np.stack([c["emb"] for c in chunks]) if chunks
else np.zeros((0, 384), dtype=np.float32))
with open(os.path.join(d, "meta.json"), "w", encoding="utf-8") as f:
json.dump(meta, f)
np.savez(os.path.join(d, "emb.npz"),
chunk_emb=chunk_emb,
visual_emb=entry["visual"]["emb"],
visual_time=entry["visual"]["time"])
except Exception as e:
if DEBUG:
print(f"[DEBUG] disk cache save failed: {e}")
def _disk_update_precise(fp: str, precise: dict) -> None:
"""Persist newly-refined precise sectors without rewriting embeddings."""
if not DISK_CACHE:
return
p = os.path.join(CACHE_DIR, fp, "meta.json")
try:
with open(p, encoding="utf-8") as f:
meta = json.load(f)
meta["precise"] = precise
with open(p, "w", encoding="utf-8") as f:
json.dump(meta, f)
except Exception:
pass
# ── Formatter ──────────────────────────────────────────────────────────────────
def format_output(regions: list[dict], prompt: str, summary: str,
mode: str = "specific") -> str:
if not regions:
return (
f' Topic not found in this video.\n\n'
f' Prompt : "{prompt}"\n\n'
f' No relevant segments found.\n'
f' Try a broader or differently worded prompt.'
)
mode_tag = {"global": "🌐 global", "hybrid": "⚡ hybrid", "specific": "🎯 specific"}.get(mode, "")
lines = [f' {len(regions)} region(s) found for: "{prompt}" [{mode_tag} · audio+video]\n']
if DEBUG:
for i, r in enumerate(regions, 1):
dur = round(r["end"] - r["start"], 1)
text = r.get("precise_text", r["text"])
ent = r.get("entity_overlap", 0)
bi = r.get("bi_score", 0)
vis = r.get("visual_score", 0)
tags = ", ".join(r.get("visual_tags", []))
lines += [
f'\n [{i}] {fmt_time(r["start"])} → {fmt_time(r["end"])} ({dur}s)',
f' {conf_bar(r["confidence"])} {r["confidence"]:.0%} '
f'(embed {bi:.0%} · entity {ent:.0%} · visual {vis:.0%})',
f' 👁 {tags}' if tags else ' 👁 (no visual tags)',
f' "{text[:220]}{"…" if len(text) > 220 else ""}"',
]
lines += [
f' 🧠 Based on uploaded video (audio + video): — "{prompt}"',
summary,
f'\n Sources: ' + ', '.join(
f'[{fmt_time(r["start"])} → {fmt_time(r["end"])}]' for r in regions
),
]
return "\n".join(lines)
# ── Cold-start ingestion (transcription + chunking + CLIP, in parallel) ───────────
def ingest_video(video_path: str) -> dict:
"""
First sight of a video: decode audio, then run the two heavy independent
stages concurrently —
• CPU thread : base transcription → semantic chunking → entities
• MPS thread : frame extraction → CLIP encoding
Returns the cache entry dict.
"""
audio = load_audio(video_path)
duration = media_duration(video_path) or (len(audio) / SAMPLE_RATE)
result = {}
def _audio_branch():
segments = transcribe_fast(audio)
result["segments"] = segments
if segments:
result["chunks"] = add_entities_to_chunks(semantic_chunk(segments))
else:
result["chunks"] = []
def _visual_branch():
result["visual"] = analyze_visual(video_path, duration)
with ThreadPoolExecutor(max_workers=2) as ex:
fa = ex.submit(_audio_branch)
fv = ex.submit(_visual_branch)
fa.result()
fv.result()
return {
"segments": result["segments"],
"chunks": result["chunks"],
"visual": result["visual"],
"audio": audio,
"precise": {},
}
# ── Global path (shared by GLOBAL mode + every escalation/fallback) ───────────────
# Global summaries run on the clean base transcript — no whole-video turbo pass.
def _run_global(chunks: list[dict], visual: dict,
prompt: str, max_words: int) -> list[dict]:
sampled = sample_chunks_for_global(chunks, _words_to_global_chunks(max_words))
print(f"[NINA] Global mode — sampled {len(sampled)} chunks evenly (base transcript).")
for c in sampled:
c.setdefault("confidence", 0.5)
c.setdefault("bi_score", 0.0)
c.setdefault("entity_overlap", 0.0)
c.setdefault("visual_score", 0.0)
c["precise_text"] = c.get("text", "") # base text is enough for an overview
return add_visual_tags(sampled, visual)
# ── Main pipeline (generator: streams progress + answer) ─────────────────────────
def format_iris(regions, prompt, best_bi):
if not regions:
return (f'No clear moments about "{prompt}" were found in this video.\n\n'
f'Try broader or differently-worded keywords.')
lines = [f'IRIS - moments matching: "{prompt}"\n']
for i, r in enumerate(sorted(regions, key=lambda x: x["start"]), 1):
txt = (r.get("precise_text", r.get("text", "")) or "").strip()
tags = ", ".join(r.get("visual_tags", []))
lines.append(f' {i}. [{fmt_time(r["start"])} -> {fmt_time(r["end"])}] {r.get("confidence", 0):.0%} match')
lines.append(f' "{txt[:260]}{"..." if len(txt) > 260 else ""}"')
if tags:
lines.append(f' on screen: {tags}')
lines.append("")
lines.append(f' {len(regions)} moment(s) - ranked by relevance, shown in order of appearance.')
return "\n".join(lines)
def run(video_path: str, prompt: str, token: str = ""):
if not _verify_usage_token(token):
yield ("This request is not authorized.\n\n"
"Please use IRIS at https://app.nnlabs.pl/iris (sign in first). "
"Direct API access is disabled, and each account has a usage limit.")
return
if video_path is None:
yield "Please upload a video."
return
if not prompt or not prompt.strip():
yield "Please type what you want to find."
return
prompt = prompt.strip()
print(f'\n[IRIS] -- Find -- "{prompt}"')
_evict_stale()
fp = None
if video_path in _cache:
entry = _cache[video_path]
else:
try: fp = _fingerprint(video_path)
except Exception: fp = None
disk = _disk_load(fp) if fp else None
if disk is not None:
disk["audio"] = load_audio(video_path)
entry = disk; _cache[video_path] = entry
else:
yield "Reading video - transcribing audio and scanning frames..."
entry = ingest_video(video_path)
if not entry["segments"]:
yield "No speech detected."
return
_cache[video_path] = entry
if fp: _disk_save(fp, entry)
entry["_atime"] = time.time()
segments, chunks, visual = entry["segments"], entry["chunks"], entry["visual"]
precise_cache, audio = entry["precise"], entry["audio"]
if fp is None:
try: fp = _fingerprint(video_path)
except Exception: fp = None
if not chunks:
yield "No speech detected."
return
candidates, best_bi = bi_encode_retrieve(chunks, prompt)
_, v_norm = visual_scores(visual, prompt)
results = rerank_and_score(candidates, prompt, visual["time"], v_norm)
regions = merge_regions(results)
if not regions:
yield format_iris([], prompt, best_bi)
return
before = dict(precise_cache)
fill_precise(audio, regions, chunks, precise_cache, RETRANSCRIBE_MAX)
if fp and precise_cache != before:
_disk_update_precise(fp, precise_cache)
regions = add_visual_tags(regions, visual)
regions = sorted(regions, key=lambda r: r.get("confidence", 0), reverse=True)[:RESULTS_TOP_N]
print(f"[IRIS] Done - {len(regions)} moment(s).")
yield format_iris(regions, prompt, best_bi)
def _synthesize_streaming(regions: list[dict], prompt: str, max_words: int, mode: str):
"""Stream the formatted answer to the UI as the LLM produces it."""
print(f"[NINA] Running synthesis [{mode}]...")
yield format_output(regions, prompt, "✍️ …", mode)
last = ""
for partial in llm_synthesize_stream(regions, prompt, max_words, mode):
last = partial
yield format_output(regions, prompt, partial, mode)
if not last:
yield format_output(regions, prompt, _fallback_summary(regions), mode)
# ── Gradio UI ──────────────────────────────────────────────────────────────────
demo = gr.Interface(
fn=run,
inputs=[
gr.Video(label="Upload video "),
gr.Textbox(
label="What moment do you want to find?",
placeholder='e.g. "What does the video say about...?" '
'(add "briefly" or "in detail" to size the answer)',
),
# Issued by app.nnlabs.pl after it counts the use; required to run.
gr.Textbox(label="access token (set automatically by the website)", value=""),
],
outputs=gr.Textbox(label="Results", lines=38),
title="IRIS v0.80 — Intelligent Retrieval & Indexing (find the moment)",
description=(
"video -> find the exact moments a topic is mentioned or shown. "
"Use it at https://app.nnlabs.pl/iris."
),
flagging_mode="never",
# Privacy: Gradio purges uploaded files older than 5 min, every 5 min. One
# upload still serves many prompts (audio/transcript are cached in RAM), and
# the file is cleaned shortly after the session goes idle.
delete_cache=(300, 300),
)
# Concurrency for ~5-10 simultaneous users on an 8-vCPU Space: run a few jobs in
# parallel, queue the rest (CTranslate2 is CPU-bound). Tune with the env var.
demo.queue(
default_concurrency_limit=int(os.getenv("NINA_CONCURRENCY", "3")),
max_size=int(os.getenv("NINA_QUEUE_MAX", "40")),
)
if __name__ == "__main__":
demo.launch()
# ─────────────────────────────────────────────
# NINA — Neural Information & Narrative Abstraction
# app.py | IRIS v0.80 | Made by Malina
# ─────────────────────────────────────────────
#
# Precision MULTIMODAL video search + summarization pipeline.
# Pipeline: video → audio → ASR segments → semantic chunking
# → bi-encoder (text) ⨉ CLIP (vision) fusion → cross-encoder + entity overlap
# → merge → confidence-gated precise re-transcribe → streaming LLM synthesis.
#
# ─────────────────────────────────────────────────────────────────────────────
# v0.80 — performance & modernisation pass over v0.71
# ─────────────────────────────────────────────────────────────────────────────
# Measured on Apple M1 (8 core / 16 GB), 39 s test clip:
# cold first prompt 58.8 s → ~9 s, warm prompts ~1.5 s → sub-second to first token.
#
# [P1] HARDWARE ACCELERATION
# Torch models now auto-target the best device (CUDA → MPS → CPU). On
# Apple Silicon CLIP runs ~1.8× and the cross-encoder ~1.4× faster on the
# Metal (MPS) backend. CTranslate2 (Whisper) has no MPS backend, so it
# stays on CPU but is tuned: int8 + cpu_threads, greedy decoding for the
# cheap pass, and Silero VAD to skip non-speech.
#
# [P2] NO MORE DOUBLE TRANSCRIPTION
# v0.71 transcribed the whole video with `base`, then RE-transcribed every
# sampled sector with the ~15× slower `large-v3-turbo` — even for a global
# summary (≈24 s wasted on this clip). v0.80:
# • GLOBAL summaries use the clean `base` transcript directly (no turbo).
# • SPECIFIC queries re-transcribe with turbo ONLY on matched regions and
# ONLY when the base pass was low-confidence (avg_logprob gate). On
# clean audio turbo never runs; on hard audio it runs exactly where it
# helps. Per-sector results are cached and persist across prompts.
#
# [P3] PARALLEL COLD START
# Audio transcription (CPU) and CLIP frame analysis (MPS) are independent
# and now run concurrently on first sight of a video, overlapping the two
# heaviest stages instead of summing them.
#
# [P4] ZERO-COPY I/O
# Audio is decoded straight from the container (PyAV) — no ffmpeg wav
# temp file. Frames are streamed from one ffmpeg rawvideo pipe into NumPy —
# no per-frame JPEG temp files.
#
# [P5] PERSISTENT DISK CACHE
# Transcript, chunk + visual embeddings and precise sectors are persisted
# under a content fingerprint, so re-opening the same video — even after a
# restart — skips transcription/chunking/CLIP entirely.
#
# [P6] STREAMING + UX
# LLM synthesis streams token-by-token; run() is a generator that shows
# progress and writes the answer as it is produced.
#
# [Q] QUALITY FIXES
# • Router no longer misreads "what does the video say about X" as a global
# summary (the bare word "about"/"explain" no longer forces global mode).
# • The illustrative "[2:30]" timestamp no longer leaks into answers.
#
# Scoring formula (SPECIFIC/HYBRID mode):
# audio = 0.6 * sigmoid(cross_encoder_logit) + 0.4 * entity_overlap
# final = (1 - VISUAL_WEIGHT) * audio + VISUAL_WEIGHT * clip_visual_score
#
# Entity extraction: regex + simple NP heuristic (no spacy dependency).
# Works multilingually because it operates on surface forms, not POS tags.
# ─────────────────────────────────────────────
import os
import re
import json
import math
import base64
import hashlib
import hmac
import shutil
import tempfile
import threading
import time
from concurrent.futures import ThreadPoolExecutor
import ffmpeg
import gradio as gr
import numpy as np
from PIL import Image
from faster_whisper import WhisperModel
from faster_whisper.audio import decode_audio
from sentence_transformers import SentenceTransformer, CrossEncoder
from huggingface_hub import InferenceClient
import warnings
warnings.filterwarnings("ignore", message="FP16 is not supported on CPU; using FP32 instead")
# ── Config ─────────────────────────────────────────────────────────────────────
WHISPER_FAST = "base"
WHISPER_PRECISE = "large-v3-turbo"
SAMPLE_RATE = 16000
WHISPER_CPU_THREADS = int(os.getenv("WHISPER_CPU_THREADS", "3"))
# 3 threads × ~3 concurrent jobs ≈ an 8-vCPU Space.
# Override via env. CTranslate2 is CPU-only.
RETRANSCRIBE_MAX = 12 # per-prompt budget for NEW (uncached) sector transcriptions