Customer Scenarios

Concrete playbooks for the most common customer situations. Each scenario walks through: what the customer asked for, which model to recommend, what hardware to size, and how to deploy.

flowchart LR
    classDef ask fill:#fff3d6,stroke:#c08800
    classDef rec fill:#e8f0ff,stroke:#3b6ad6
    classDef deploy fill:#d9f5e0,stroke:#1f8f3a

    Q[Customer asks
something] --> SCN{Which scenario?}
    SCN --> S1[Semantic search
in offline ES]:::ask
    SCN --> S2[Replacing
OpenAI embeddings]:::ask
    SCN --> S3[Multilingual
support]:::ask
    SCN --> S4[Image
search]:::ask
    SCN --> S5[Code
search]:::ask
    SCN --> S6[Better
reranking]:::ask

    S1 --> M1[jina-embeddings-v5-text-small
+ jina-reranker-v3]:::rec
    S2 --> M2[jina-embeddings-v5-text-nano
drop-in]:::rec
    S3 --> M3[jina-embeddings-v5-text-small
89 languages]:::rec
    S4 --> M4[jina-clip-v2
text and image]:::rec
    S5 --> M5[jina-code-embeddings-1.5b]:::rec
    S6 --> M6[jina-reranker-v3
131K context]:::rec

Loading

Scenario A - Offline semantic search in Elasticsearch

Customer: a bank with an on-prem ES cluster, indexing internal research notes. No outbound internet.

Recommend: jina-embeddings-v5-text-small (multilingual, 1024-dim) for indexing + jina-reranker-v3 for top-K reranking. Both fit on a single L4.

Hardware: 1x NVIDIA L4 (24GB VRAM) per host, 100GB disk for both bundles + cache. Two replicas behind a load balancer if QPS > 50.

Deploy:

1. On a connected machine, run ./scripts/pull-prebuilt.sh jina-embeddings-v5-text-small gpu and ./scripts/pull-prebuilt.sh jina-reranker-v3 gpu. Output: two .tar.gz files.
2. Transfer both to the air-gapped host (SCP, USB, or whatever the bank's change-control allows).
3. docker load < ... and docker run --gpus all -p 8080:8080 ... for each on different ports.
4. Wire ES inference service to both:

PUT _inference/text_embedding/jina-embed
{"service": "openai", "service_settings": {
  "url": "http://embed-host:8080/v1/embeddings",
  "model_id": "jina-embeddings-v5-text-small",
  "api_key": "not-needed"
}}

PUT _inference/rerank/jina-rerank
{"service": "cohere", "service_settings": {
  "url": "http://rerank-host:8081/v1/rerank",
  "model_id": "jina-reranker-v3",
  "api_key": "not-needed"
}}

5. Index documents through the embedding inference and use the reranker on top-50 retrieval.

Scenario B - Drop-in replacement for OpenAI embeddings

Customer: a SaaS company whose paying customers refuse to send their data to OpenAI.

Recommend: jina-embeddings-v5-text-nano (CPU-friendly, 768-dim). The smallest model that's still a credible quality replacement.

Hardware: CPU is fine for < 100 QPS - 4 vCPU + 4GB RAM. Add a GPU only if latency is critical.

Deploy: 5 lines of bash:

docker load < jina-embeddings-v5-text-nano-cpu.tar.gz
docker run -d -p 8080:8080 jina/jina-embeddings-v5-text-nano:cpu

Client change - just rebase the OpenAI SDK URL:

from openai import OpenAI
client = OpenAI(base_url="http://your-host:8080/v1", api_key="not-needed")
client.embeddings.create(model="jina-embeddings-v5-text-nano", input=["Hello"])

That's it. No application code change beyond base_url.

Scenario C - Multilingual support across 89 languages

Customer: a global support-ticket triage system that must work in English, German, Spanish, Chinese, Japanese, Arabic.

Recommend: jina-embeddings-v5-text-small. Trained on 89 languages with task-aware routing.

Sizing: ~3GB VRAM, fits anywhere. L4 or T4 if GPU.

Tip: use the task field. retrieval for semantic search, classification for the triage classifier, clustering for grouping similar tickets. See API Reference -> Embedding tasks.

Scenario D - Image and text search

Customer: a fashion retailer that wants visual + text search ("red dress like this image, but in cotton").

Recommend: jina-clip-v2 (multimodal embeddings, text and image in the same space).

Hardware: ~4GB VRAM. L4 minimum if you want < 100ms image embedding.

API call (image + text fused into a single embedding):

curl http://host:8080/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{"input": [{"content": [
    {"type": "text", "text": "cotton dress"},
    {"type": "image", "format": "base64", "value": "<BASE64>"}
  ]}]}'

Image must be base64-encoded inline (no URLs - the air-gap forbids outbound fetches). Max 10 MB per input.

Scenario E - Code search

Customer: dev tools company indexing 10M code files for repo search.

Recommend: jina-code-embeddings-1.5b (code-aware, 1536-dim) or 0.5b for lighter deploys.

Hardware: 1.5b needs ~4GB VRAM. 0.5b runs on CPU for low-volume.

Tip: build per-language indexes, then query with task: "retrieval". Code-specific reranking with jina-reranker-v3 on top of the top-100.

Scenario F - Better reranking on top of existing retrieval

Customer: already has a retrieval pipeline (BM25 or any embedding model), wants to add a strong reranker.

Recommend: jina-reranker-v3 (131K context, multilingual, top-quality).

Hardware: ~3GB VRAM. CPU works for top-K < 50, GPU for higher volumes.

Deploy alongside existing system, no changes to retrieval:

curl http://rerank-host:8080/v1/rerank \
  -H "Content-Type: application/json" \
  -d '{
    "query": "best vector database for embeddings",
    "documents": ["...top 50 from your retriever..."],
    "top_n": 10
  }'

131K context means you can pass long documents (full pages, transcripts) without chunking.

Scenario G - Hospital with HIPAA + sovereign data

Customer: hospital with patient records, must run in their data center, no AI vendor allowed.

Recommend: jina-embeddings-v5-text-nano or text-small. The smallest workable model lowers the audit/security review burden.

Compliance pitch:

• "Model weights and dependencies are on customer hardware. We do not see your data or your queries."
• "Container has HF_HUB_OFFLINE=1 and TRANSFORMERS_OFFLINE=1 set - no path exists for the model to fetch anything at runtime."
• "Image is built once on a connected machine, transferred under your change-control process, run on offline machines indefinitely."
• "No license server, no phone-home, no telemetry."
• "Source code is on GitHub (Apache-2.0 for the CLI), reviewable by your security team."

Pair this with Why Air-Gap -> What "air-gap" means in this project when responding to the customer's security questionnaire.

Next

• Picking a Model - decision tree for model selection
• Sizing & Hardware - capacity planning per model
• Quick Start - hands-on first-deploy walkthrough