Multimodal Embedding & Reranker Models with Sentence Transformers

Sentence Transformers is a Python library for using and training embedding and reranker models for applications like retrieval augmented generation, semantic search, and more. With the v5.4 update, you can now encode and compare texts, images, audio, and videos using the same familiar API. In this blogpost, I'll show you how to use these new multimodal capabilities for both embedding and reranking.

Multimodal embedding models map inputs from different modalities into a shared embedding space, while multimodal reranker models score the relevance of mixed-modality pairs. This opens up use cases like visual document retrieval, cross-modal search, and multimodal RAG pipelines.

Table of Contents

• What are Multimodal Models?
• Installation
• Multimodal Embedding Models
• Loading a Model
• Encoding Images
• Cross-Modal Similarity
• Encoding Queries and Documents
• Multimodal Reranker Models
• Ranking Mixed-Modality Documents
• Predicting Pair Scores
• Retrieve and Rerank
• Input Formats and Configuration
• Supported Input Types
• Checking Modality Support
• Processor and Model kwargs
• Supported Models
• Additional Resources

What are Multimodal Models?

Traditional embedding models convert text into fixed-size vectors. Multimodal embedding models extend this by mapping inputs from different modalities (text, images, audio, or video) into a shared embedding space. This means you can compare a text query against image documents (or vice versa) using the same similarity functions you're already familiar with.

Similarly, traditional reranker (Cross Encoder) models compute relevance scores between pairs of texts. Multimodal rerankers can score pairs where one or both elements are images, combined text-image documents, or other modalities.

For example, you can compare a text query against image documents, find video clips matching a description, or build RAG pipelines that work across modalities.

Installation

Multimodal models require some extra dependencies. Install the extras for the modalities you need (see Installation for more details):

# For image support
pip install -U "sentence-transformers[image]"

# For audio support
pip install -U "sentence-transformers"

# For video support
pip install -U "sentence-transformers"

# Mix and match as needed
pip install -U "sentence-transformers[image,video,train]"

VLM-based models like Qwen3-VL-2B require a GPU with at least ~8 GB of VRAM. For the 8B variants, expect ~20 GB. If you don't have a local GPU, consider using a cloud GPU service or Google Colab. On CPU, these models will be extremely slow; text-only or CLIP models are better suited for CPU inference.

Multimodal Embedding Models

Loading a Model

Loading a multimodal embedding model works exactly like loading a text-only model:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Qwen/Qwen3-VL-Embedding-2B", revision="refs/pr/23")

The revision argument is required for now because the integration pull requests for these models are still pending. Once they're merged, you'll be able to load them without specifying a revision.

The model automatically detects which modalities it supports, so there's nothing extra to configure. See Processor and Model kwargs if you want to control things like image resolution or model precision.

Encoding Images

With a multimodal model loaded, model.encode() accepts images alongside text. Images can be provided as URLs, local file paths, or PIL Image objects (see Supported Input Types for all accepted formats):

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Qwen/Qwen3-VL-Embedding-2B", revision="refs/pr/23")

# Encode images from URLs
img_embeddings = model.encode([
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg",
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
])
print(img_embeddings.shape)
# (2, 2048)

Cross-Modal Similarity

You can compute similarities between text embeddings and image embeddings, since the model maps both into the same space:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Qwen/Qwen3-VL-Embedding-2B", revision="refs/pr/23")

# Encode images
img_embeddings = model.encode([
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg",
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
])

# Encode text queries (one matching + one hard negative per image)
text_embeddings = model.encode([
 "A green car parked in front of a yellow building",
 "A red car driving on a highway",
 "A bee on a pink flower",
 "A wasp on a wooden table",
])

# Compute cross-modal similarities
similarities = model.similarity(text_embeddings, img_embeddings)
print(similarities)
# tensor([[0.5115, 0.1078],
# [0.1999, 0.1108],
# [0.1255, 0.6749],
# [0.1283, 0.2704]])

As expected, "A green car parked in front of a yellow building" is most similar to the car image (0.51), and "A bee on a pink flower" is most similar to the bee image (0.67). The hard negatives ("A red car driving on a highway", "A wasp on a wooden table") correctly receive lower scores.

You might notice that even the best matching scores (0.51, 0.67) aren't very close to 1.0. This is due to the modality gap: embeddings from different modalities tend to cluster in separate regions of the space. Cross-modal similarities are typically lower than within-modal ones (e.g., text-to-text), but the relative ordering is preserved, so retrieval still works well.

Encoding Queries and Documents

For retrieval tasks, encode_query() and encode_document() are the recommended methods. Many retrieval models prepend different instruction prompts depending on whether the input is a query or a document, similar to how chat models might apply different system prompts depending on the goal. Model authors can specify their prompts in the model config, and encode_query() / encode_document() automatically load and apply the correct one:

• encode_query() uses the model's "query" prompt (if available) and sets task="query".
• encode_document() uses the first available prompt from "document", "passage", or "corpus", and sets task="document".

Under the hood, both are thin wrappers around encode(), they just handle prompt selection for you. Here's what cross-modal retrieval looks like:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Qwen/Qwen3-VL-Embedding-2B", revision="refs/pr/23")

# Encode text queries with the query prompt
query_embeddings = model.encode_query([
 "Find me a photo of a vehicle parked near a building",
 "Show me an image of a pollinating insect",
])

# Encode document screenshots with the document prompt
doc_embeddings = model.encode_document([
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg",
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
])

# Compute similarities
similarities = model.similarity(query_embeddings, doc_embeddings)
print(similarities)
# tensor([[0.3907, 0.1490],
# [0.1235, 0.4872]])

These methods accept the same input types as encode() (images, URLs, multimodal dicts, etc.) and pass through the same parameters. For models without specialized query/document prompts, they behave identically to encode().

Multimodal Reranker Models

Multimodal reranker (CrossEncoder) models score the relevance between pairs of inputs, where each element can be text, an image, audio, video, or a combination. They tend to outperform embedding models in terms of quality, but are slower since they process each pair individually. The currently available pretrained multimodal rerankers focus on text and image inputs, but the architecture supports any modality that the underlying model can handle.

Ranking Mixed-Modality Documents

The rank() method scores and ranks a list of documents against a query, supporting mixed modalities:

from sentence_transformers import CrossEncoder

model = CrossEncoder("Qwen/Qwen3-VL-Reranker-2B", revision="refs/pr/11")

query = "A green car parked in front of a yellow building"
documents = [
 # Image documents (URL or local file path)
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg",
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
 # Text document
 "A vintage Volkswagen Beetle painted in bright green sits in a driveway.",
 # Combined text + image document
 {
 "text": "A car in a European city",
 "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg",
 },
]

rankings = model.rank(query, documents)
for rank in rankings:
 print(f"{rank['score']:.4f}\t(document {rank['corpus_id']})")
"""
0.9375 (document 0)
0.5000 (document 3)
-1.2500 (document 2)
-2.4375 (document 1)
"""

The reranker correctly identifies the car image (document 0) as the most relevant result, followed by the combined text+image document about a car in a European city (document 3). The bee image (document 1) scores lowest.

Keep in mind that the modality gap can influence absolute scores: text-image pair scores may occupy a different range than text-text or image-image pair scores.

You can also check which modalities a reranker supports using modalities and supports(), just like with embedding models:

print(model.modalities)
# ['text', 'image', 'video', 'message']

print(model.supports("image"))
# True

# Check if the model supports a specific pair of modalities
print(model.supports(("image", "text")))
# True

Predicting Pair Scores

You can also use predict() to get raw relevance scores for specific pairs of inputs:

from sentence_transformers import CrossEncoder

model = CrossEncoder("jinaai/jina-reranker-m0", trust_remote_code=True)

scores = model.predict([
 ("A green car", "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg"),
 ("A bee on a flower", "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"),
 ("A green car", "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"),
])
print(scores)
# [0.9389156 0.96922314 0.46063158]

Retrieve and Rerank

A common pattern is to use an embedding model for fast initial retrieval, then refine the top results with a reranker:

from sentence_transformers import SentenceTransformer, CrossEncoder

# Step 1: Retrieve with an embedding model
embedder = SentenceTransformer("Qwen/Qwen3-VL-Embedding-2B", revision="refs/pr/23")

query = "revenue growth chart"
query_embedding = embedder.encode_query(query)

# Pre-compute corpus embeddings (do this once, then store them)
document_screenshots = [
 "path/to/doc1.png",
 "path/to/doc2.png",
 # ... potentially millions of document screenshots
]
corpus_embeddings = embedder.encode_document(document_screenshots, show_progress_bar=True)

# Simple cosine similarity retrieval, viable as long as embeddings fit in memory
similarities = embedder.similarity(query_embedding, corpus_embeddings)
top_k_indices = similarities.argsort(descending=True)[0][:10]

# Step 2: Rerank the top-k results with a reranker model
reranker = CrossEncoder("nvidia/llama-nemotron-rerank-vl-1b-v2", trust_remote_code=True)

top_k_documents = [document_screenshots[i] for i in top_k_indices]
rankings = reranker.rank(query, top_k_documents)
for rank in rankings:
 print(f"{rank['score']:.4f}\t{top_k_documents[rank['corpus_id']]}")

Since the corpus embeddings are pre-computed, the initial retrieval is fast even over millions of documents. The reranker then provides more accurate scoring over the smaller candidate set.

Input Formats and Configuration

Supported Input Types

Multimodal models accept a variety of input formats. Here's a summary of what you can pass to model.encode():

Checking Modality Support

You can check which modalities a model supports using the modalities property and supports() method:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Qwen/Qwen3-VL-Embedding-2B", revision="refs/pr/23")

# List all supported modalities
print(model.modalities)
# ['text', 'image', 'video', 'message']

# Check for a specific modality
print(model.supports("image"))
# True
print(model.supports("audio"))
# False

The "message" modality indicates that the model accepts chat-style message inputs with interleaved content. In practice, you rarely need to use this directly. When you pass strings, URLs, or multimodal dicts, the model converts them to the appropriate message format internally. Sentence Transformers supports two message formats:

1. Structured (most VLMs, e.g. Qwen3-VL): Content is a list of typed dicts, e.g. [{"type": "text", "text": "..."}, {"type": "image", "image": ...}]
2. Flat (e.g. Deepseek-V3): Content is a direct value, e.g. "some text"

The format is auto-detected from the model's chat template.

Since all inputs get converted into the same message format internally, you can mix input types in a single encode() call:

embeddings = model.encode([
 # A text input
 "A green car parked in front of a yellow building",
 # An image input (URL)
 "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg",
 # A combined text + image input
 {
 "text": "A car in a European city",
 "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg",
 },
])

Processor and Model kwargs

You may want to control image resolution bounds or model precision. Use processor_kwargs and model_kwargs when loading the model:

model = SentenceTransformer(
 "Qwen/Qwen3-VL-Embedding-2B",
 model_kwargs={"attn_implementation": "flash_attention_2", "torch_dtype": "bfloat16"},
 processor_kwargs={"min_pixels": 28 * 28, "max_pixels": 600 * 600},
 revision="refs/pr/23",
)

processor_kwargs controls how inputs are preprocessed (e.g., image resolution bounds). Higher max_pixels means higher quality but more memory and compute. These are

The post Multimodal Embedding & Reranker Models with Sentence Transformers first appeared on Hard Wired.

Context Engineering: Nová disciplína, která mění pravidla AI

Jak jsem se naučil, že úspěch AI aplikací nezávisí na dokonalém promptu, ale na tom, co model "vidí" kolem něj

Před třemi lety jsem trávil hodiny ladění promptů. Psal jsem stránkové instrukce, experimentoval s různými formulacemi, testoval desítky variant. A přesto můj AI asistent zapomínal klíčové informace z předchozích konverzací, můj kódovací pomocník ztrácel přehled o architektuře projektu a RAG systém nedokázal propojit souvislosti napříč dokumenty.

Pak jsem pochopil zásadní věc: problém nebyl v tom, jak jsem se modelu ptal, ale v tom, co všechno model věděl v okamžiku, kdy odpovídal. Objevil jsem context engineering – disciplínu, která překračuje hranice prompt engineeringu a mění celou hru.

Proč prompt engineering přestal stačit

Když poprvé otevřete ChatGPT, připadá vám to jednoduché: napíšete otázku, dostanete odpověď. Jenže reality produkčních AI aplikací je jiná. Představte si, že stavíte AI asistenta pro zákaznický servis. Potřebuje:

• Znát historii všech předchozích interakcí s klientem
• Mít přístup k aktuálním informacím o produktech
• Rozumět firemním procesům a pravidlům
• Pamatovat si kontext celé konverzace
• Umět zavolat externí API pro ověření dat

Žádný prompt, ať je sebevíc dokonalý, to sám nezvládne. Potřebujete systém, který modelu poskytne správný kontext ve správný čas. To je podstata context engineeringu.

Co je context engineering v praxi

Context engineering je disciplína navrhování a budování systémů, které orchestrují všechny informace, nástroje a paměť potřebné k tomu, aby AI dokázala řešit složité, real-world úkoly.

Nejde jen o prompt. Jde o celý informační ekosystém kolem modelu.

Reálný příklad z mé praxe

Nedávno jsem stavěl AI asistenta pro právní kancelář. Klasický přístup by byl:

Jsi právní expert. Odpovídej na otázky klientů o smluvním právu.

Context engineering přístup vypadal takto:

1. Systémový kontext:

Role: Senior právní poradce specializující se na obchodní právo
Firma: [název], 15 let praxe, focus na SaaS a tech startupy  
Regulatory environment: České právo, EU regulace

2. Dynamický retrieval:

# Při každé otázce systém:
query = user_question
relevant_cases = vector_search(query, case_database)
current_legislation = api_call("legal_updates", query)
client_history = get_client_context(client_id)
firm_templates = search_templates(query)

3. Paměťový systém:

# Kontext se skládal z:
- Dlouhodobé paměti klienta (preference, předchozí případy)
- Krátkodobé paměti konverzace (co už probrali dnes)
- Faktual knowledge base (zákony, judikáty)
- Tool access (kalkulačky poplatků, termíny soudů)
- Meta-context (urgence, složitost případu)

Výsledek? Místo obecných právních rad model poskytoval konkrétní doporučení založená na historii klienta, aktuální legislativě a firemních postupech.

Anatomie profesionálního context systému

1. Multi-layer memory architecture

Immediate context - co model "vidí" právě teď:

• Aktuální prompt a konverzace
• Výsledky z právě provedených nástrojů
• Dočasný stav úkolu

Session memory - co si pamatuje během práce:

• Historie kroků a rozhodnutí
• Předchozí výsledky a chyby
• Evoluce strategie řešení

Long-term memory - trvalé znalosti:

• User profily a preference
• Learnt patterns a insights
• Firemní knowledge base

2. Intelligent retrieval orchestration

Nejsložitější část. Systém musí v real-time rozhodnout:

• Které dokumenty jsou relevantní
• Jaké externí API zavolat
• Kolik kontextu použít (token limits)
• V jakém pořadí informace poskytovat

Můj workflow:

def build_context(user_query, session_state):
    # 1. Analýza query
    intent = classify_intent(user_query)
    entities = extract_entities(user_query)

    # 2. Multi-source retrieval
    docs = semantic_search(user_query, weight=0.4)
    tools = suggest_tools(intent, weight=0.3) 
    memory = get_relevant_memory(session_state, weight=0.3)

    # 3. Context assembly
    context = assemble_context(
        system_prompt=get_system_prompt(intent),
        retrieved_docs=docs[:5],  # Top 5 to stay within limits
        available_tools=tools,
        conversation_memory=memory,
        user_profile=get_user_context()
    )

    return context

3. Dynamic context optimization

Context není statický. Mění se podle:

Task complexity - složité úkoly potřebují víc kontextu
User expertise - expert vs. beginner potřebuje jiné informace
Performance feedback - učení se z úspěchů a chyb
Resource constraints - tokens, latency, costs

Praktické techniky z praxe

Context Layering

Místo jednoho obřího promptu stavím kontext po vrstvách:

# Layer 1: Core identity
system_role = """
Senior business analyst s 10+ lety zkušeností
Specializace: SaaS metriky, customer analytics
Styl: Data-driven, konkrétní doporučení
"""

# Layer 2: Current task context  
task_context = f"""
Aktuální projekt: {project_name}
Deadline: {deadline}
Stakeholders: {stakeholder_list}
Previous insights: {session_memory}
"""

# Layer 3: Dynamic information
dynamic_context = f"""
Relevantní data: {retrieved_data}
Dostupné nástroje: {available_tools}
Aktuální metrics: {live_metrics}
"""

Context Chaining

Pro komplexní úkoly rozdělím práci do kroků, kde výstup jednoho kroku se stává kontextem pro další:

# Krok 1: Analýza problému
problem_analysis = llm_call(
    context=base_context + user_problem,
    task="Analyzuj problém a identifikuj klíčové otázky"
)

# Krok 2: Sběr dat s kontextem z kroku 1
data_context = base_context + problem_analysis
retrieved_data = gather_data(problem_analysis.key_questions)

# Krok 3: Řešení s full kontextem
solution = llm_call(
    context=data_context + retrieved_data,
    task="Navrhni řešení založené na analýze a datech"
)

Adaptive Context Compression

Když se blížím k token limitu, používám kompresní strategie:

def compress_context(context_items, max_tokens):
    if calculate_tokens(context_items) <= max_tokens:
        return context_items

    # Prioritizace podle důležitosti
    prioritized = rank_by_relevance(context_items)

    # Postupná komprese
    compressed = []
    token_budget = max_tokens

    for item in prioritized:
        if item.type == "critical":
            compressed.append(item)  # Vždy zahrnout
        elif item.type == "supporting":
            if token_budget > estimate_tokens(item):
                compressed.append(summarize(item))  # Komprese

    return compressed

Časté problémy a jejich řešení

Context Poisoning

Problém: Chyba se dostane do kontextu a pak se propaguje dál.

Řešení z praxe:

def validate_context(context_item):
    # Fact-checking pro kritické informace
    if context_item.type == "factual":
        confidence = fact_check(context_item.content)
        if confidence < 0.8:
            context_item.add_disclaimer("Unverified information")

    # Timestamp check pro časově citlivé info
    if context_item.age > MAX_STALENESS:
        refresh_data(context_item)

    return context_item

Context Overload

Problém: Příliš mnoho informací rozptyluje model.

Mé řešení:

• Používám "attention hints" - explicitně říkám, na co se zaměřit
• Strukturuji kontext hierarchicky (nejdůležitější nahoře)
• Implementuji "context budgeting" - každý typ info má limit

context_budget = {
    "system_instructions": 500,   # tokens
    "user_input": 1000,
    "retrieved_docs": 2000,
    "tool_outputs": 1500,
    "memory": 1000
}

Tool Confusion

Problém: Model si vybírá špatné nástroje.

Moje strategie:

def smart_tool_selection(user_intent, available_tools):
    # Jen relevantní nástroje pro daný typ úkolu
    if user_intent == "data_analysis":
        return [tools.python_executor, tools.data_visualizer]
    elif user_intent == "web_research":
        return [tools.web_search, tools.summarizer]

    # Nikdy nedávat všechny nástroje najednou
    return filter_tools_by_relevance(available_tools, max_count=5)

Frameworky a nástroje

LangChain/LangGraph

Skvělé pro orchestraci workflows, ale pozor na over-engineering:

from langgraph import StateGraph

# Definuji workflow s explicitním context flow
workflow = StateGraph()
workflow.add_node("analyze", analyze_with_context)
workflow.add_node("retrieve", smart_retrieval) 
workflow.add_node("synthesize", synthesize_response)

# Context se propaguje mezi kroky
workflow.add_edge("analyze", "retrieve")
workflow.add_edge("retrieve", "synthesize")

LlamaIndex

Exceluje v knowledge management:

from llama_index import VectorStoreIndex, ContextBuilder

# Automatické budování kontextu
context_builder = ContextBuilder()
context_builder.add_memory_layer(user_profile)
context_builder.add_retrieval_layer(document_index)
context_builder.add_tool_layer(available_functions)

Anthropic MCP

Nejnovější standard pro propojení AI s externí systémy:

# MCP server pro firemní data
mcp_server = MCPServer()
mcp_server.register_resource("customer_db", CustomerDatabase())
mcp_server.register_tool("send_email", EmailTool())

# AI má strukturovaný přístup k firemním systémům

Budoucnost context engineeringu

Vidím tři hlavní trendy:

1. Automated Context Assembly
AI začíná samo rozpoznávat, jaký kontext potřebuje. Experiments s "self-reflective agents" ukazují zajímavé výsledky.

2. Multi-Modal Context Integration
Kombinace textu, obrázků, audio, video do jednotného kontextu. Pracuji na projektu, kde AI analyzuje video cally a extrahuje kontext pro další rozhodnutí.

3. Collaborative Context Networks
Více AI agentů sdílí kontext a buduje kolektivní "paměť" týmu.

Co si odnést domů

Context engineering není jen technická disciplína – je to nový způsob myšlení o AI aplikacích. Moje klíčová doporučení:

1. Začněte s auditem kontextu
Podívejte se na vaše současné AI aplikace. Co všechno model "nevidí", ale měl by?

2. Investujte do memory systémů
Dlouhodobá paměť je game-changer. AI, které si pamatuje vaše preference a zkušenosti, je kvalitativně jiné.

3. Experimentujte s context compression
Naučte se čistit a komprimovat kontext. Méně může být více.

4. Měřte context effectiveness
Trackujte, které části kontextu model skutečně používá. Optimalizujte na základě dat.

5. Myslĕte systémově
Context engineering je systémová disciplína. Nejde o izolované prompty, ale o architekturu informačních toků.

A především: context engineering je budoucnost AI aplikací. Kdo ho zvládne dřív, získá obrovskou výhodu.

Po několika letech experimentování s LLM si myslím, že context engineering je nejdůležitější skill pro AI builders. Není to jen o tom dát modelu správné informace – je to o pochopení toho, jak AI "myslí" a jak navrhnout systémy, které s tímto myšlením spolupracují. Je to fascinující kombinace software architecture, cognitive science a trochy magie.

The post Context Engineering: Nová disciplína, která mění pravidla AI first appeared on Hard Wired.