Sync docs/ from main repo
f5ab06c
Sync regenerated On-Prem CLI screenshots
a8a4cb4
Sync docs/ from main repo (rename to Jina On-Prem)
572600e
Sync docs/ from main repo
f7685a3
Sync docs/ from main repo
5299c4c
Sync docs/ from main repo
f90d0d0
Sync docs/ from main repo
Adds Product & Model Lifecycle (EOL) page and its nav links; also picks up
pending docs/ updates that had not yet been mirrored to the wiki.
defbe6f
Add Product & Model Lifecycle (EOL) policy page
Defines support/maintenance/EOL for self-managed air-gapped Jina models.
Key point: SM models are never force-retired (customer holds weights);
EOL governs the maintenance term, not access. 18mo-from-GA or 12mo-from-
replacement-GA maintenance term (whichever later), advisory notice
periods, current per-generation lifecycle status. Linked in sidebar.
16afa9c
Sizing-And-Hardware: :gpu-opt-fast -> :gpu-opt max-throughput env flags; fix multi-task default; add lowest-L4 95k
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
f128c22
docs(sizing): correct nano fast to the STANDARD single-L4 box (g2-standard-8)
nano:gpu-opt-fast hits 102,195 tok/s end-to-end on the standard g2-standard-8
(1 L4, 8 vCPU, single process) — no vCPU upgrade — after eliminating the double
tokenization (tokenize once, reuse ids). Replaces the earlier g2-standard-32
figure. cos_min 0.9999988 vs fp32.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
35d74eb
docs(sizing): small :gpu-opt-fast — 131,704 tok/s on 8xL4 + quality clarity
small clears 100k via horizontal scaling: 8 replicas on one g2-standard-96
(8xL4), 131,704 tok/s aggregate. Add the one-replica-per-GPU recipe. Clarify
quality: no 16-bit small clears cos>=0.9999 vs fp32 (a 28-layer precision floor)
and neither does the stock :gpu small (~0.99988 both); :gpu-opt-fast small is
cos>=0.9999 vs the stock server = zero regression vs the deployed model.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
84d68c4
docs(sizing): add :gpu-opt-fast (nano 100k+ tok/s) frontier section
Document the merge + torch.compile(emulate_precision_casts) + lean-pipeline
stack: nano hits 105,806 tok/s end-to-end on one L4 (g2-standard-32) at cos_min
0.9999987 vs fp32. Note the single-task + high-vCPU caveats, and that -small is
FLOPs-capped (~16-20k on one L4; needs replicas/bigger GPU for 100k).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
75a97ae
docs(sizing): correct co-located replica figure (~1.25x, knee at 2)
Measured 2-replica-per-L4 aggregate is ~1.25x in the concurrent regime (not the
earlier ~1.5x estimate, pre-base64); 3 replicas regress (GPU over-subscription).
Also note flash-attn-2 unpadding gives no gain here (length-sorting already
removes the padding it would target).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
55b735b
docs(sizing): clarify pure-encode vs end-to-end throughput
The "Throughput math" ~30k tok/s figure is raw model.encode() at bs32 (no
server/concurrency) — an upper bound, not what a client sees. Add a callout
distinguishing it from the :gpu-opt end-to-end numbers (27k-36k), note the
measured ~44k bs128 kernel ceiling, and mark client-batching as only needed on
the stock images (:gpu-opt batches server-side).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2f05027
docs(sizing): :gpu-opt now fp16 + base64; update L4 throughput numbers
fp16 default (matches stock :gpu, cos-sim 0.9999981 vs fp32, batch-invariant;
bf16 headroom was thin and load-dependent). Document encoding_format=base64 as
the throughput path (OpenAI SDK default, byte-identical) and refresh the
end-to-end numbers: nano c64 ~27k, bulk ~36k tok/s; small c64 ~8.7k, bulk ~11k.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
e997880
docs(sizing): document :gpu-opt dynamic-batching images + L4 throughput
Add a "GPU dynamic batching" section to Sizing & Hardware covering the
:gpu-opt image tags (server-side request coalescing, bf16 default), the
measured end-to-end L4 throughput (5-9x on concurrent small requests),
the new JINA_BATCH_TOKENS / JINA_ATTN / JINA_BATCH_WAIT_MS / JINA_MAX_BATCH_ROWS
knobs, and a Quick-Start pointer.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
b01209e
Document CPU bf16 autocast + runtime tuning env vars
- Sizing-And-Hardware: note the AMX/AVX512_BF16 ~2x CPU speedup, add a
Runtime tuning section documenting JINA_CPU_AUTOCAST, OMP_NUM_THREADS,
JINA_DTYPE (none were documented before)
- FAQ + Comparison: one-line mentions of CPU bf16 autocast
42b3592
Sync docs/ from main repo
3ef9744
Sync docs/ from main repo
2a0c3b9
visual overhaul: pixel art + ASCII + real terminal recordings
Hero illustrations (replacing 5 heavy isometric '3D banger' images with pixel art):
- pixel-bundle-flow.png -> Home (cloud + floppy + server)
- pixel-rocket.png -> Quick-Start (rocket + motion lines)
- pixel-airgap-perimeter.png -> Why-Airgap (2 servers in dotted boundary)
- pixel-mascot-shield.png -> Architecture (cute robot with shield)
- removed 5 hero-*.png (too 3D, too gradient-heavy)
ASCII architecture diagrams (10 mermaid blocks now plain code-block ASCII):
- Home: at-a-glance flow
- Why-Airgap: SaaS vs jina-airgap comparison
- Architecture: two-phase model, Docker image layers, multi-schema server
- Sizing-And-Hardware: deployment topologies, redundancy patterns
- Bundling-Guide: bundle flow
- API-Reference: schema routing
- Cheat-Sheet: two phases
Mermaid kept where ASCII would be too messy (sequence diagrams, decision trees).
Real terminal recordings (light theme, no fake headers):
- 01-list, 02-bundle: real CLI sessions
- 03-deploy, 04-schemas: container pre-started OUTSIDE recording so the first
frame shows the first curl, not a blank load-wait
- Recorded with locally-built image carrying the SHORT_MODEL_ID fix
- Dropped SVGs (CSS-animated, don't render statically; GIFs cover the same)
6c4d95a
Sync docs/ from main repo
466611d
Add Cheat-Sheet, Glossary, Comparison hero illustration
Cheat-Sheet: one-page reference for SAs to screenshot on customer calls.
Pitch, commands, headline models, 4 schemas, hardware rules, common errors,
the 3 sentences that close most calls. Reachable from sidebar top section.
Glossary: 30+ terms defined for non-technical readers - air-gap, bundle,
ColBERT, matryoshka, omni, task, VRAM, etc. Reachable under Reference.
Comparison hero: 3 roads branching to SaaS/server/shielded-vault icons.
Reinforces the decision framing of the page.
a2afea9
Add hero illustrations to Why-Airgap + Quick-Start, link k8s manifest
Z-Image Turbo hero illustrations:
- Why-Airgap: data center inside a glowing security perimeter dome with cloud
icons blocked outside. Visual metaphor for what air-gap means.
- Quick-Start: stopwatch + running docker whale, reinforcing the 5-min promise.
Sizing-And-Hardware: replaced inline k8s YAML stub with a link to the new
k8s/jina-airgap.yaml manifest (Namespace + Deployment + Service + HPA + Ingress).
d84fc0a
Add Z-Image Turbo hero illustrations to Home and Architecture
Home: hero-bundle-flow.png shows the 3-phase flow as isometric infographic
(cloud download -> USB transfer -> shielded server). Kept the 03-deploy.gif
below as the demo.
Architecture: hero-server-shielded.png shows a server rack inside a glowing
shield bubble - thematic match for the architecture page.
Generated via authenticated gradio call to mrfakename/Z-Image-Turbo Space
(720x1280).
93d3924
Regenerate visuals + add static PNGs + more diagrams
GIFs/SVGs:
- New demo scripts (no char-by-char typing animation)
- First frame now shows informative title banner instead of blank
- agg --font-size 18 --theme github-dark --last-frame-duration 3
- 02-bundle now shows the --dry-run output (much cleaner than full build)
- 03/04 re-recorded with locally-built image carrying the SHORT_MODEL_ID fix
so PNGs show clean 'jina-embeddings-v5-text-nano' (not 'jinaai/...')
New static PNGs (last frame of each GIF) for places where animation isn't available.
Wiki diagrams added:
- Why-Airgap: request-lifecycle sequence diagram showing zero outbound calls
- Sizing-And-Hardware: deployment topology diagram (sidecar / shared / behind ES)
74d264c
Add Comparison and Versioning-And-Updates pages
Comparison: jina-airgap vs Ollama / vLLM / ONNX / hosted api.jina.ai / DIY
Versioning-And-Updates: zero-downtime rollouts, bundle pinning, reindex
strategy, customer-side change control
Both linked from Home and the regrouped _Sidebar.
a9de2bd
Update --network=none note - CLI no longer prints the misleading hint
0c00656
Redesign wiki for SA/sales audience
- Add Why-Airgap explaining the concept + comparison to SaaS/VPC endpoints
- Add Customer-Scenarios with 7 industry-specific playbooks
- Add Picking-A-Model decision tree
- Add Sizing-And-Hardware with capacity planning + k8s example
- Add Architecture with multi-diagram explanation
- Add FAQ covering business / technical / sales-objection questions
- Rewrite Home, Quick-Start, Bundling-Guide, API-Reference, Troubleshooting
- Add _Footer.md for persistent nav
- Regrouped _Sidebar.md by audience (start / playbooks / deploy / reference)
- Heavy use of mermaid diagrams for non-technical readers
- Documented Cohere response shape, GHCR auth requirement, --network=none gotcha
a611e90
Sync docs/ from main repo
7c1510d