import spaces # must be imported before torch on ZeroGPU import base64 import html import io import os import time import av import gradio as gr import librosa import numpy as np import torch from PIL import Image from sentence_transformers import SentenceTransformer MODEL_ID = "google/embeddinggemma-2" HERE = os.path.dirname(os.path.abspath(__file__)) ASSETS = os.path.join(HERE, "assets") # bfloat16 on GPU (float16 overflows this model). Loaded once at module scope. model = SentenceTransformer(MODEL_ID, device="cuda", model_kwargs={"dtype": torch.bfloat16}) model.max_seq_length = 4096 DIMS = [768, 512, 256, 128] TEXT_TASKS = { "Web search": "SearchQuery", "Question answering": "QuestionAnswering", "Fact checking": "FactChecking", "Code search": "CodeRetrieval", } SYMMETRIC_TASKS = { "Sentence similarity": "SentenceSimilarity", "Classification": "Classification", "Clustering": "Clustering", } MAX_VIDEO_FRAMES = 24 SAMPLE_IMAGES = [ ("cat.jpg", "Pallas's cat walking in snow"), ("cats_couch.jpg", "Two cats asleep on a pink couch"), ("bee.jpg", "Bee on a pink cosmos flower"), ("baklava.jpg", "Baklava on a plate"), ("mosque.jpg", "Domed mosque with minarets"), ("squirrel.jpg", "Cubist squirrel painting"), ("tiger.jpg", "Screenshot: classifier says 'tiger'"), ] SAMPLE_AUDIO = [ ("jfk.wav", "JFK speech clip"), ("mlk.flac", "MLK speech clip"), ("kungfu.mp3", "Movie line: 'I know kung fu'"), ("musicgen.wav", "Short generated music clip"), ] SAMPLE_TEXTS = """A wild cat with thick fur prowls through a snowy forest. Bees pollinate flowers while collecting nectar. Layers of filo pastry, chopped nuts and honey syrup. Ask not what your country can do for you. The orchestra played a lively upbeat tune. Abstract geometric art inspired by Picasso. Tourists gather outside a historic building in Istanbul.""" DEFAULT_DOCS = """Northern lights | Auroras happen when charged particles from the sun collide with gases high in Earth's atmosphere. Ocean currents | Large-scale currents move heat around the planet and shape regional climates. Eiffel Tower | The wrought-iron tower was built in Paris for the 1889 World's Fair. Sleep | During deep sleep the brain consolidates memories formed during the day. Photosynthesis | Plants convert sunlight, water and carbon dioxide into glucose and oxygen. Auroras (ES) | Las auroras boreales se producen cuando el viento solar choca con la atmósfera. binary_search.py | def search(xs, t):\\n lo, hi = 0, len(xs) - 1\\n while lo <= hi:\\n mid = (lo + hi) // 2 Markets | The stock market closed higher on Friday after strong earnings reports.""" DEFAULT_SENTENCES = """The cat is sleeping on the sofa. Le chat dort sur le canapé. Die Katze schläft auf dem Sofa. A dog is barking in the garden. Quarterly revenue grew by 12 percent. Los ingresos trimestrales crecieron un 12 por ciento.""" # ----------------------------------------------------------------------------- helpers def _load_audio(path: str) -> dict: """Decode any audio file to 16 kHz mono float32, as the model expects.""" wav, _ = librosa.load(path, sr=16000, mono=True, duration=300) return {"array": wav.astype(np.float32), "sampling_rate": 16000} def _load_video(path: str) -> list[Image.Image]: """Sample a video at ~1 fps (the model's default) into a list of PIL frames.""" frames: list[Image.Image] = [] with av.open(path) as container: stream = container.streams.video[0] fps = float(stream.average_rate or 25) step = max(1, round(fps)) for i, frame in enumerate(container.decode(stream)): if i % step == 0: img = frame.to_image().convert("RGB") img.thumbnail((640, 640)) frames.append(img) if len(frames) >= MAX_VIDEO_FRAMES: break if not frames: raise gr.Error("Could not read any frames from that video.") return frames def _to_model_input(item: dict): kind, value = item["kind"], item["value"] if kind == "text": return value if kind == "image": return {"image": Image.open(value).convert("RGB")} if kind == "audio": return {"audio": _load_audio(value)} if kind == "video": return {"video": _load_video(value)} raise ValueError(kind) def _encode(items: list[dict], prompt_name: str | None, dim: int) -> np.ndarray: """Embed a mixed list of items. Task prefixes apply to text only; media gets none.""" out = np.zeros((len(items), int(dim)), dtype=np.float32) text_idx = [i for i, it in enumerate(items) if it["kind"] == "text"] media_idx = [i for i, it in enumerate(items) if it["kind"] != "text"] if text_idx: vecs = model.encode( [items[i]["value"] for i in text_idx], prompt_name=prompt_name, truncate_dim=int(dim), normalize_embeddings=True, convert_to_numpy=True, batch_size=16, show_progress_bar=False, ) out[text_idx] = vecs.astype(np.float32) # Media is encoded one at a time: modalities and audio lengths differ per item. for i in media_idx: vec = model.encode( [_to_model_input(items[i])], truncate_dim=int(dim), normalize_embeddings=True, convert_to_numpy=True, show_progress_bar=False, )[0] out[i] = vec.astype(np.float32) return out def _thumb_b64(path: str, size: int = 120) -> str: img = Image.open(path).convert("RGB") img.thumbnail((size, size)) buf = io.BytesIO() img.save(buf, format="JPEG", quality=80) return "data:image/jpeg;base64," + base64.b64encode(buf.getvalue()).decode() def _label(path: str, fallback: str | None = None) -> str: return fallback or os.path.basename(path) def _gallery_paths(gallery) -> list[tuple[str, str]]: out = [] for item in gallery or []: img, cap = item, None if isinstance(item, (list, tuple)): img, cap = item[0], (item[1] if len(item) > 1 else None) if isinstance(img, dict): img = img.get("path") or img.get("name") if isinstance(img, str) and os.path.isfile(img): out.append((img, _label(img, cap))) return out def _file_paths(files) -> list[str]: out = [] for f in files or []: p = f if isinstance(f, str) else getattr(f, "name", None) or (f.get("path") if isinstance(f, dict) else None) if p and os.path.isfile(p): out.append(p) return out ICONS = {"text": "📝", "image": "🖼️", "audio": "🔊", "video": "🎬"} def _results_html(ranked: list[tuple[dict, float]], note: str) -> str: rows = [] top = max((s for _, s in ranked), default=1.0) or 1.0 for rank, (item, score) in enumerate(ranked, start=1): if item["kind"] == "image": preview = f'' else: preview = f'
{ICONS[item["kind"]]}
' width = max(2.0, 100.0 * max(score, 0.0) / top) rows.append( f"""
{rank}
{preview}
{item['kind']}{html.escape(item['label'])}
{score:.3f}
""" ) return f'
{html.escape(note)}
{"".join(rows)}
' # ----------------------------------------------------------------------------- tab 1: any-to-any search def _build_corpus(texts: str, gallery, audio_files) -> list[dict]: corpus: list[dict] = [] for line in (texts or "").splitlines(): if line.strip(): corpus.append({"kind": "text", "value": line.strip(), "label": line.strip()}) for path, label in _gallery_paths(gallery): corpus.append({"kind": "image", "value": path, "label": label}) audio_labels = {os.path.join(ASSETS, f): lbl for f, lbl in SAMPLE_AUDIO} for path in _file_paths(audio_files): base = os.path.basename(path) label = next((lbl for f, lbl in SAMPLE_AUDIO if base == f), None) corpus.append({"kind": "audio", "value": path, "label": audio_labels.get(path, label or base)}) return corpus @spaces.GPU(duration=90) def cross_modal_search( query_type: str, query_text: str, query_image: str | None, query_audio: str | None, query_video: str | None, corpus_texts: str, corpus_images: list | None, corpus_audio: list | None, dim: int = 768, ) -> str: """Search a mixed corpus of text, images and audio with a text, image, audio or video query. Every item is embedded into EmbeddingGemma 2's single shared vector space, truncated to `dim` (Matryoshka) and L2-normalized, then ranked by cosine similarity. Args: query_type: One of "Text", "Image", "Audio", "Video". query_text: Query string (used when query_type is "Text"). query_image: Path to a query image (used when query_type is "Image"). query_audio: Path to a query audio file (used when query_type is "Audio"). query_video: Path to a query video file (used when query_type is "Video"). corpus_texts: Text items in the corpus, one per line. corpus_images: Images in the corpus. corpus_audio: Audio files in the corpus. dim: Embedding dimension: 768, 512, 256 or 128. Returns: HTML listing the corpus ranked by similarity to the query. """ t0 = time.perf_counter() dim = int(dim) if query_type == "Text": if not (query_text or "").strip(): raise gr.Error("Type a text query.") query = {"kind": "text", "value": query_text.strip(), "label": query_text.strip()} elif query_type == "Image": if not query_image: raise gr.Error("Upload a query image.") query = {"kind": "image", "value": query_image, "label": "query image"} elif query_type == "Audio": if not query_audio: raise gr.Error("Upload or record query audio.") query = {"kind": "audio", "value": query_audio, "label": "query audio"} else: if not query_video: raise gr.Error("Upload a query video.") query = {"kind": "video", "value": query_video, "label": "query video"} corpus = _build_corpus(corpus_texts, corpus_images, corpus_audio) if not corpus: raise gr.Error("The corpus is empty — add some text, images or audio.") if len(corpus) > 48: raise gr.Error("Keep the corpus to 48 items or fewer.") # Text queries use the search-query prefix; text corpus items use the document format. q_vec = _encode([query], "SearchQuery", dim)[0] docs = [ dict(it, value=f"title: none | text: {it['value']}") if it["kind"] == "text" else it for it in corpus ] c_vecs = _encode(docs, None, dim) scores = np.clip(c_vecs @ q_vec, -1.0, 1.0) # bf16 can nudge self-matches past 1 order = np.argsort(-scores) ranked = [(corpus[i], float(scores[i])) for i in order] counts = {k: sum(1 for c in corpus if c["kind"] == k) for k in ("text", "image", "audio")} note = ( f"{query_type} query vs {counts['text']} texts · {counts['image']} images · {counts['audio']} audio clips" f" · {dim}-d · {time.perf_counter() - t0:.1f}s" ) return _results_html(ranked, note) # ----------------------------------------------------------------------------- tab 2: text retrieval + matryoshka def _parse_docs(raw: str) -> list[tuple[str, str]]: docs = [] for line in (raw or "").splitlines(): line = line.strip() if not line: continue title, content = ("none", line) if " | " in line: title, content = (s.strip() for s in line.split(" | ", 1)) content = content.replace("\\n", "\n") if content: docs.append((title or "none", content)) return docs @spaces.GPU(duration=45) def text_search(query: str, documents: str, task: str = "Web search") -> list[list]: """Rank documents for a query at every Matryoshka size (768, 512, 256, 128 dims). The query gets the task prefix for the chosen task; each document is formatted as `title: {title} | text: {content}`. Shows how ranking holds up as vectors shrink. Args: query: The search query. documents: One document per line, optionally `Title | text`. task: One of "Web search", "Question answering", "Fact checking", "Code search". Returns: Rows of [title, document, cosine@768, rank@768, rank@512, rank@256, rank@128]. """ query = (query or "").strip() docs = _parse_docs(documents) if not query: raise gr.Error("Enter a query.") if not docs: raise gr.Error("Add at least one document, one per line.") if len(docs) > 40: raise gr.Error("Use at most 40 documents.") prompt = TEXT_TASKS[task] # Embed once at full size, then truncate + re-normalize for each Matryoshka size. q = model.encode(query, prompt_name=prompt, convert_to_numpy=True, show_progress_bar=False).astype(np.float32) d = model.encode( [f"title: {t} | text: {c}" for t, c in docs], convert_to_numpy=True, show_progress_bar=False ).astype(np.float32) ranks, full_scores = {}, None for dim in DIMS: qd = q[:dim] / np.linalg.norm(q[:dim]) dd = d[:, :dim] / np.linalg.norm(d[:, :dim], axis=1, keepdims=True) s = np.clip(dd @ qd, -1.0, 1.0) if dim == 768: full_scores = s order = np.argsort(-s) r = np.empty(len(docs), dtype=int) r[order] = np.arange(1, len(docs) + 1) ranks[dim] = r rows = [] for i in np.argsort(-full_scores): title, content = docs[i] snippet = content if len(content) <= 140 else content[:137] + "…" rows.append([title, snippet, round(float(full_scores[i]), 4)] + [int(ranks[d_][i]) for d_ in DIMS]) return rows # ----------------------------------------------------------------------------- tab 3: similarity matrix def _heatmap_html(labels: list[str], sim: np.ndarray) -> str: n = len(labels) lo, hi = float(sim[~np.eye(n, dtype=bool)].min()) if n > 1 else 0.0, 1.0 head = "".join(f'#{j + 1}' for j, l in enumerate(labels)) body = [] for i, l in enumerate(labels): cells = [] for j in range(n): v = float(sim[i, j]) t = 0.0 if hi == lo else (v - lo) / (hi - lo) alpha = 0.08 + 0.82 * max(0.0, min(1.0, t)) cells.append(f'{v:.2f}') body.append(f'#{i + 1} {html.escape(l[:60])}{"".join(cells)}') return f'
{head}{"".join(body)}
' @spaces.GPU(duration=45) def similarity_matrix(sentences: str, task: str = "Sentence similarity", dim: int = 768) -> str: """Pairwise cosine similarity between lines of text (one per line), as a heatmap. Every line gets the same symmetric task prefix. Works across languages. Args: sentences: Two to sixteen lines of text. task: One of "Sentence similarity", "Classification", "Clustering". dim: Embedding dimension: 768, 512, 256 or 128. Returns: HTML heatmap of the similarity matrix. """ lines = [s.strip() for s in (sentences or "").splitlines() if s.strip()] if len(lines) < 2: raise gr.Error("Enter at least two lines.") if len(lines) > 16: raise gr.Error("Use at most 16 lines.") vecs = _encode([{"kind": "text", "value": s} for s in lines], SYMMETRIC_TASKS[task], int(dim)) return _heatmap_html(lines, np.clip(vecs @ vecs.T, -1.0, 1.0)) # ----------------------------------------------------------------------------- tab 4: raw embedding (API / MCP) @spaces.GPU(duration=30) def embed(text: str, prompt_name: str = "SearchQuery", dim: int = 768) -> list[float]: """Return the EmbeddingGemma 2 embedding for a piece of text. Args: text: Text to embed. prompt_name: Task prefix name, e.g. SearchQuery, Document, QuestionAnswering, FactChecking, CodeRetrieval, Classification, Clustering, SentenceSimilarity. dim: Output dimension: 768, 512, 256 or 128 (L2-normalized). Returns: The embedding as a list of floats. """ if not (text or "").strip(): raise gr.Error("Enter some text.") if prompt_name not in model.prompts: raise gr.Error(f"Unknown prompt_name. Choose from: {', '.join(sorted(model.prompts))}") v = model.encode( text, prompt_name=prompt_name, truncate_dim=int(dim), normalize_embeddings=True, convert_to_numpy=True ) return [round(float(x), 6) for x in v] # ----------------------------------------------------------------------------- UI CSS = """ .gradio-container { max-width: 1180px !important; margin: 0 auto !important; } .dark .gradio-container { color: var(--body-text-color); } .eg-note { font-size: .85rem; opacity: .75; margin: 2px 0 10px; } .eg-list { display: flex; flex-direction: column; gap: 6px; } .eg-row { display: grid; grid-template-columns: 28px 64px 1fr 64px; align-items: center; gap: 10px; padding: 6px 10px; border: 1px solid var(--border-color-primary); border-radius: 10px; background: var(--block-background-fill); } .eg-row:first-child { border-color: var(--color-accent); } .eg-rank { font-weight: 700; text-align: center; opacity: .7; } .eg-prev img { width: 64px; height: 48px; object-fit: cover; border-radius: 6px; display: block; } .eg-icon { width: 64px; height: 48px; display: grid; place-items: center; font-size: 22px; border-radius: 6px; background: var(--background-fill-secondary); } .eg-label { font-size: .92rem; line-height: 1.3; } .eg-kind { font-size: .7rem; text-transform: uppercase; letter-spacing: .04em; opacity: .6; margin-right: 8px; } .eg-bar { height: 6px; border-radius: 3px; background: var(--background-fill-secondary); margin-top: 6px; overflow: hidden; } .eg-bar span { display: block; height: 100%; background: var(--color-accent); border-radius: 3px; } .eg-score { font-variant-numeric: tabular-nums; text-align: right; font-weight: 600; } .eg-heat { overflow-x: auto; } .eg-heat table { border-collapse: collapse; font-size: .82rem; } .eg-heat th, .eg-heat td { padding: 6px 8px; text-align: center; border: 1px solid var(--border-color-primary); } .eg-heat td { font-variant-numeric: tabular-nums; min-width: 44px; } .eg-heat .eg-rowlab { text-align: left; font-weight: 500; white-space: nowrap; } """ INTRO = f"""# EmbeddingGemma 2 — one vector space for text, images, audio & video [{MODEL_ID}](https://huggingface.co/{MODEL_ID}) is Google DeepMind's 740M multimodal embedding model: 768-d vectors, 100+ languages, task prefixes, and Matryoshka truncation to 512 / 256 / 128 dims. """ with gr.Blocks(title="EmbeddingGemma 2 Explorer") as demo: gr.Markdown(INTRO) with gr.Tabs(): # ---------------------------------------------------------------- any-to-any with gr.Tab("🔀 Any-to-any search"): gr.Markdown( "Query with **text, an image, audio or a video** and rank a mixed corpus of text, images and audio. " "Everything lives in the same embedding space." ) with gr.Row(): with gr.Column(scale=5): q_type = gr.Radio(["Text", "Image", "Audio", "Video"], value="Text", label="Query type") q_text = gr.Textbox( label="Text query", value="a fluffy animal in the snow", lines=2, visible=True ) q_image = gr.Image(label="Image query", type="filepath", height=220, visible=False) q_audio = gr.Audio(label="Audio query", type="filepath", visible=False) q_video = gr.Video(label="Video query (sampled at 1 fps, ≤24 frames)", visible=False) dim1 = gr.Radio(DIMS, value=768, label="Dimensions (Matryoshka)") go1 = gr.Button("Search", variant="primary") with gr.Accordion("Corpus", open=False): c_texts = gr.Textbox(label="Texts (one per line)", value=SAMPLE_TEXTS, lines=8) c_images = gr.Gallery( label="Images", value=[(os.path.join(ASSETS, f), c) for f, c in SAMPLE_IMAGES], type="filepath", columns=4, height=260, interactive=True, ) c_audio = gr.File( label="Audio clips", value=[os.path.join(ASSETS, f) for f, _ in SAMPLE_AUDIO], file_count="multiple", file_types=["audio"], ) with gr.Column(scale=6): out1 = gr.HTML(label="Ranked results") def _toggle(kind): return [gr.update(visible=kind == k) for k in ("Text", "Image", "Audio", "Video")] q_type.change(_toggle, q_type, [q_text, q_image, q_audio, q_video], api_visibility="private") inputs1 = [q_type, q_text, q_image, q_audio, q_video, c_texts, c_images, c_audio, dim1] go1.click(cross_modal_search, inputs1, out1, api_name="cross_modal_search") q_text.submit(cross_modal_search, inputs1, out1, api_visibility="private") gr.Examples( label="Try", examples=[ ["Text", "a fluffy animal in the snow", None, None], ["Text", "something sweet to eat", None, None], ["Text", "a famous political speech", None, None], ["Text", "un insecto sobre una flor", None, None], ["Text", "a user interface showing model predictions", None, None], ["Image", "", os.path.join(ASSETS, "cats_couch.jpg"), None], ["Audio", "", None, os.path.join(ASSETS, "jfk.wav")], ], inputs=[q_type, q_text, q_image, q_audio], ) # ---------------------------------------------------------------- text retrieval with gr.Tab("📚 Text retrieval & Matryoshka"): gr.Markdown( "Task-prefixed retrieval. The table shows each document's rank at **768 / 512 / 256 / 128** dims " "— a quick look at how much you can shrink vectors before ranking shifts." ) with gr.Row(): with gr.Column(scale=5): q2 = gr.Textbox(label="Query", value="What causes the northern lights?", lines=2) task2 = gr.Dropdown(list(TEXT_TASKS), value="Web search", label="Task prefix") docs2 = gr.Textbox( label="Documents", info="One per line, optional `Title | text`.", value=DEFAULT_DOCS, lines=12 ) go2 = gr.Button("Rank", variant="primary") with gr.Column(scale=7): out2 = gr.Dataframe( headers=["Title", "Document", "Cosine @768", "Rank @768", "@512", "@256", "@128"], datatype=["str", "str", "number", "number", "number", "number", "number"], interactive=False, wrap=True, label="Results", ) go2.click(text_search, [q2, docs2, task2], out2, api_name="text_search") q2.submit(text_search, [q2, docs2, task2], out2, api_visibility="private") gr.Examples( label="Try", examples=[ ["What causes the northern lights?", DEFAULT_DOCS, "Web search"], ["When was the Eiffel Tower built?", DEFAULT_DOCS, "Question answering"], ["Solar wind creates auroras.", DEFAULT_DOCS, "Fact checking"], ["find an item in a sorted list", DEFAULT_DOCS, "Code search"], ], inputs=[q2, docs2, task2], outputs=out2, fn=text_search, cache_examples=True, cache_mode="lazy", ) # ---------------------------------------------------------------- similarity with gr.Tab("🌐 Similarity matrix"): gr.Markdown("Pairwise cosine similarity — note how translations cluster across languages.") with gr.Row(): with gr.Column(scale=4): s3 = gr.Textbox(label="Lines", value=DEFAULT_SENTENCES, lines=10) task3 = gr.Dropdown(list(SYMMETRIC_TASKS), value="Sentence similarity", label="Task prefix") dim3 = gr.Radio(DIMS, value=768, label="Dimensions") go3 = gr.Button("Compare", variant="primary") with gr.Column(scale=6): out3 = gr.HTML() go3.click(similarity_matrix, [s3, task3, dim3], out3, api_name="similarity_matrix") gr.Examples( label="Try", examples=[ [DEFAULT_SENTENCES, "Sentence similarity", 768], [ "I love this phone, the battery lasts forever.\nThe screen cracked after one day.\n" "Ce téléphone est génial.\nPantalla rota, muy decepcionado.\nBest purchase I made this year.", "Classification", 256, ], ], inputs=[s3, task3, dim3], outputs=out3, fn=similarity_matrix, cache_examples=True, cache_mode="lazy", ) # ---------------------------------------------------------------- raw embedding with gr.Tab("🧮 Embed"): gr.Markdown("Get the raw vector — also exposed via the API and as an MCP tool.") with gr.Row(): with gr.Column(): t4 = gr.Textbox(label="Text", value="What causes the northern lights?", lines=3) p4 = gr.Dropdown(sorted(model.prompts), value="SearchQuery", label="prompt_name") d4 = gr.Radio(DIMS, value=256, label="Dimensions") go4 = gr.Button("Embed", variant="primary") with gr.Column(): out4 = gr.JSON(label="Embedding") go4.click(embed, [t4, p4, d4], out4, api_name="embed") gr.Markdown( "Runs in bfloat16 on ZeroGPU. Text uses the model's task prefixes; images, audio (16 kHz mono) and video " "(1 fps) are embedded without a prefix. Sample media from Hugging Face documentation datasets." ) if __name__ == "__main__": demo.launch(theme=gr.themes.Citrus(), css=CSS, mcp_server=True)