On-Prem Version History
This page lists all LLMWhisperer on-prem releases in reverse chronological order. Add newer versions at the top.
Download Configuration Files
The sample.onprem.values.yaml file — and the multiaz.values.yaml overlay used for multi-AZ deployments — for any release are bundled inside the Helm chart. To extract them, ensure you have authenticated with the Helm registry, then run:
VERSION=<VERSION> && \
TMP_DIR="/tmp/llmw-extract-$(date +%s)" && \
mkdir -p "$TMP_DIR" && \
helm pull oci://us-central1-docker.pkg.dev/pandoras-tamer/charts/llmwhisperer \
--version $VERSION --untar --untardir "$TMP_DIR" && \
cp "$TMP_DIR/llmwhisperer/sample.onprem.values.yaml" "./sample.onprem.values-$VERSION.yaml" && \
cp "$TMP_DIR/llmwhisperer/resource-values/multiaz.values.yaml" "./multiaz.values-$VERSION.yaml" && \
rm -rf "$TMP_DIR"
Releases before v2.64.4 also bundle an onprem-profile.values.yaml that must
be passed alongside onprem.values.yaml. To extract all three files — handling
the fact that multiaz.values.yaml is absent before v2.57.2 — use this script
instead:
VERSION=<VERSION> && \
TMP_DIR="/tmp/llmw-extract-$(date +%s)" && \
mkdir -p "$TMP_DIR" && \
helm pull oci://us-central1-docker.pkg.dev/pandoras-tamer/charts/llmwhisperer \
--version $VERSION --untar --untardir "$TMP_DIR" && \
cp "$TMP_DIR/llmwhisperer/sample.onprem.values.yaml" "./sample.onprem.values-$VERSION.yaml" && \
cp "$TMP_DIR/llmwhisperer/resource-values/onprem-profile.values.yaml" "./onprem-profile.values-$VERSION.yaml" && \
{ cp "$TMP_DIR/llmwhisperer/resource-values/multiaz.values.yaml" "./multiaz.values-$VERSION.yaml" 2>/dev/null || true; }
rm -rf "$TMP_DIR"
The onprem-profile.values.yaml copy uses && (it must succeed), while the
multiaz.values.yaml copy uses || true (it is genuinely optional — absent
before v2.57.2).
Replace <VERSION> with the target release version (e.g., v2.56.4).
v2.67.1 — October 27, 2026
New Features
include_cell_bordersfor Excel extraction — choose whether the workbook's own cell borders are drawn into the extracted text. Defaults totrue, so existing output is unchanged; set it tofalseto drop them and save tokens downstream. See Cell borders in Excel extraction.
Fixes & Improvements
- PgDog connection pooler — all database traffic now goes through PgDog, which pools connections in front of PostgreSQL. It is on by default and needs action before upgrading; see Upgrade Notes.
- The low-cost mode worker now runs Tesseract in-process; the separate
tesseract-ocrsidecar andtesseractimage are retired. - HA RabbitMQ brokers now add missing quorum-queue members on their own.
- Excel: General-format numbers now render as Excel displays them, sheets with a long styled-but-empty tail are extracted correctly, and a sheet that can't be read now fails with a clear error instead of silently.
- Passwords containing
/in a managed Redis or RabbitMQ URL are now redacted in the chart's validation messages.
Upgrade Notes
-
PgDog connection pooler — action required. PgDog is deployed by default and reads the database settings you already have in
global.sharedConfigs.database(or itsexistingSecret), but it cannot readglobal.*placement, so it needs its own copy of your node selector, tolerations, and image pull secret. With a values file made from an earlier sample,helm upgradestops before applying anything, with a message naming the values to set. Before upgrading:- Copy the
pgdog:block from the bottom of the newsample.onprem.values.yamlinto youronprem.values.yaml. - Change
global.imagePullSecrets:toimagePullSecrets: &globalImagePullSecrets(thepgdog:block aliases it, as it does&nsKey0/&nsVal0and&globalTolerations). If you added extraglobal.nodeSelectorpairs, add a matching*nsKeyN : *nsValNline underpgdog.nodeSelectortoo. - If you pull from your own registry, also mirror
pgdog:v0.1.60andmultiaz-migrate:v1.0.0fromus-central1-docker.pkg.dev/pandoras-tamer/unstract-llm-whisperer-on-prem/, and addimage: {repository: <your-registry>/pgdog}to thepgdog:block.
The upgrade restarts every app pod once, and pods that use the database wait for PgDog before they start.
- Copy the
-
Multi-AZ overlay adds PgDog HA —
multiaz.values.yamlruns three PgDog replicas spread across the zones that have eligible nodes (one per zone on a three-zone cluster), with thellmwhisperer-criticalPriorityClass and a PodDisruptionBudget keeping two available. Pool sizes are per replica, so PostgreSQL sees up to 3 ×pgdog.defaultPoolSizeserver connections in steady state and briefly 4 × during a rollout; leave room for that, plus any other database clients, in yourmax_connections. -
Tesseract sidecar removed from
worker-page-extraction-low-cost. Delete anysidecar:block under that worker, andTESSERACT_API_URLif you set it; you no longer need to mirror thetesseractimage. The worker's CPU and memory now scale with--concurrency(2.3 CPU / 3Gi at the default of 2, up from 1.1 CPU for the old worker plus sidecar), so make sure your nodes have room. If your values set this worker's CPU or memory resources, the upgrade stops and lists those keys: delete them, or setworker-page-extraction-low-cost.slotSizing.setResources: falseand size the worker yourself to cover Tesseract too. -
HA RabbitMQ only (
rabbitmq.ha.enabled: true) — check before upgrading. Quorum-queue membership reconciliation (rabbitmq.ha.membershipReconciliation, on by default) rewritesrabbitmq.conf, so the upgrade restarts each broker once, one at a time. Workers reconnect on their own, but tasks running over a dropped connection may run again, so upgrade in a quiet window. A broker will not stop while it holds a quorum queue that would lose quorum without it, so a queue with fewer members than brokers can hang the restart for days. Fix those first:# List quorum queues and their members (add -p <vhost> if your broker URL names one)
kubectl -n <namespace> exec rabbitmq-server-0 -c rabbitmq -- \
rabbitmqctl list_queues name type members
# For each broker missing from a queue's members, run once:
kubectl -n <namespace> exec rabbitmq-server-0 -c rabbitmq -- \
rabbitmq-queues grow rabbit@rabbitmq-server-<n>.rabbitmq-nodes.<namespace> allDo not force-delete a broker pod stuck in
Terminating; run the steps above instead. To keep the previous behaviour and skip the broker restarts, setrabbitmq.ha.membershipReconciliation: false.
You can extract and diff the config files for this release against your own values using the Download Configuration Files steps above.
v2.65.0 — October 08, 2026
New Features
- Managed Redis and RabbitMQ — LLMWhisperer can now use an external / managed Redis (including TLS) and an external / managed RabbitMQ broker (including
amqps://brokers signed by a private CA) in place of the bundled ones. See Using a Managed Redis and RabbitMQ. - The whisper-detail API now returns the
document_name, and the PDF-to-images status response shape has been updated.
Fixes & Improvements
- Tables with merged header cells now keep a rectangular RST grid.
- Hardened the application's Redis clients against dead connections.
- Disabled Celery gossip and mingle on the workers, cutting broker chatter.
- Added an index on whisper hash (built online, without blocking writes) to speed up whisper lookups.
- Richer diagnostic logs: the effective whisper parameters, OCR engine identity, and per-page render vs. line-detection timings.
v2.64.4 — August 20, 2026
Fixes & Improvements
- Automated single-AZ → multi-AZ migration as opt-in Helm upgrade hooks, plus a high-priority PriorityClass (
llmwhisperer-critical) for the stateful pods so the scheduler can preempt lower-priority workloads instead of leaving themPendingunder capacity pressure. - Accept the correctly-spelled
page_separatorquery parameter; the misspelledpage_seperatoris deprecated. - Lowered the default
word_confidence_thresholdto0.05. - Guard whisper status updates against backward transitions.
- Default the OpenTelemetry exporter protocol to
http/protobuf(fork-safe under the Celery prefork workers). - Fixed on-prem worker image resolution so a nulled worker key no longer degrades the pod to a stale backend image.
- Confined Redis/Sentinel NetworkPolicy ingress to their own namespace, and applied a Redis image CVE fix.
Upgrade Notes
onprem-profile.values.yamlis no longer used. Its resource and scaling defaults are now built into the chart, so on-prem deployments use a singleonprem.values.yaml(plus themultiaz.values.yamloverlay for multi-AZ). If your previous (v2.62.0) deploy passed-f onprem-profile.values.yaml, drop that file from your upgrade command.- Delete any bare worker keys from your
onprem.values.yamlwhen you drop the profile. The v2.62.0 sample carried four keys with no value —worker-ingestion,worker-document-collation,worker-result-webhook,libreoffice— thatonprem-profile.values.yamlused to fill in. With the profile gone, a bare (null) key discards that workload's chart defaults and the pod resolves to a stalebackend:v2.0.0image (no Celery args, broken LibreOffice conversion,ImagePullBackOffon air-gapped mirrors). Remove those four lines — the chart now supplies their defaults — or reconcile against the v2.64.4sample.onprem.values.yaml, which no longer contains them. - OpenTelemetry tracing now defaults to
http/protobuf, and the endpoint example uses the HTTP port (4318). No action unless you enabled tracing with an explicitgrpcprotocol/endpoint. - Multi-AZ overlay updated —
multiaz.values.yamlnow enables the Redis volume reset (multiazMigration.allowRedisVolumeReset: true) that the migration uses, and addsminDomainsguidance for a true 3-zone spread. The offline queue cutover stays opt-in: you setrabbitmq.ha.migrateQueues=trueyourself for the single upgrade that runs it. See the Single-AZ to Multi-AZ Migration guide.
You can extract and diff the config files for this release against your own values using the Download Configuration Files steps above.
v2.62.0 — July 06, 2026
New Features
- OpenTelemetry tracing support for on-prem: opt-in distributed tracing across the backend and all workers, emitting spans for HTTP requests, Celery tasks, database, Redis, and outbound HTTP calls to any OTLP-compatible collector. Disabled by default.
Upgrade Notes
- New optional OpenTelemetry tracing block in
sample.onprem.values.yaml(underglobal.backend.config, applying to the backend and all workers). Tracing is off by default (ENABLE_TRACING: "false"), so no action is required for a standard upgrade. To enable it, setENABLE_TRACING: "true"andOTEL_TRACES_EXPORTER: "otlp", and pointOTEL_EXPORTER_OTLP_ENDPOINT(required when tracing is enabled) at your OTLP collector. AdditionalOTEL_*options control the transport protocol, TLS, service name, excluded URLs, and trace sampling. See Observability & Tracing (OpenTelemetry) for the full setup. - The on-prem resource profile (
onprem-profile.values.yaml) now launches the backend and workers directly viauv/celery/pythoninstead of the previous shell-wrapped binary, to support the tracing instrumentation. No action is required unless you overrode thecommand/argsfor these workloads in your own values — in that case, re-sync them from the new sample file.
You can extract and diff the config files for this release against your own values using the Download Configuration Files steps above.
v2.61.0 — July 01, 2026
Fixes & Improvements
- Security patch addressing critical and high severity CVEs across all workload container images
v2.60.10 — June 29, 2026
Fixes & Improvements
- HA hardening: opt-in PriorityClass for critical stateful components (Redis, RabbitMQ) lets the scheduler preempt lower-priority workloads instead of leaving them Pending
- HA hardening: opt-in Redis Sentinel cold-start guard overlay to make Redis startup more resilient
- Broker reliability: broker publish retries are now logged, and pyamqp socket tuning is honored
- Bounded Redis result-backend timeouts to stop the chord pub/sub wedge that could leave task chains stuck
- Helm chart:
global.nodeSelectoris now a list so per-service and RabbitMQ node selectors no longer double up; zot node selector/tolerations now follow the global placement - Helm chart: image pull secrets are now configurable through a single global knob
- Removed the bundled chart image-listing scripts in favor of the docs-hosted unified
list-onprem-images.sh
Upgrade Notes
global.nodeSelectoris now a list instead of a map (action required if you customized node placement). Node placement is set once underglobal.nodeSelectoras a list of- key:/value:pairs and is inherited by all workloads; the old per-servicenodeSelectorentries are no longer needed. If you previously set node placement, move it into the singleglobal.nodeSelectorlist using the new list format.- New
global.tolerationslist — set this once if your node pool is tainted; it applies to all workloads. No action needed if your nodes aren't tainted. - New
global.imagePullSecrets— registry pull secrets are now set in one place. Backward-compatible: it defaults to theartifact-registrysecret existing installs already used, so no action is needed for a standard upgrade. If your registry needs no pull secret (no-auth/anonymous or node-IAM), setglobal.imagePullSecrets: [], or set your own docker-registry secret name. - (multi-AZ overlay) New opt-in
priorityClass—resource-values/multiaz.values.yamlcan now create allmwhisperer-criticalPriorityClass for Redis and RabbitMQ so the scheduler can preempt lower-priority pods instead of leaving HA components Pending in a capacity-constrained AZ. No action unless you run multi-AZ HA.
You can extract and diff the config files for this release against your own values using the Download Configuration Files steps above.
v2.60.8 — June 19, 2026
Fixes & Improvements
- HTML uploads now render via headless Chrome instead of LibreOffice, improving table positioning and page geometry for Word/Outlook "Save as Web Page" exports, and keeping nested tables from being split across PDF page breaks (doc/ppt/xls conversions continue to use LibreOffice)
- On-prem image registry override is now configured in a single place — set
global.imageto point all images at a private or air-gapped registry instead of editing multiple locations; per-serviceimage.registryoverrides still take precedence, and the olderglobal.imageDefaults/image-overrides.yamlapproach is deprecated
Upgrade Notes
- A new
global.image.registryfield insample.onprem.values.yamlis the single place to point all images at a private or air-gapped registry. No action is needed if you use Unstract's default registry; per-serviceimage.registryoverrides still take precedence, and the olderglobal.imageDefaults/image-overrides.yamlapproach is deprecated. - If you change
global.image.registryto your own registry, you must also setglobal.zot.enabled: false. The Zot pull-through cache only supports Unstract's default registry, and the install will stop with an error otherwise.
v2.60.7 — June 17, 2026
Fixes & Improvements
- Excel/XLSX handling hardened: style lookups now guard against missing or out-of-range references, and render dimensions are capped to prevent worker out-of-memory errors
- TesseractOCR: improved recovery of table column boundaries and removed a duplicate error from overlay output
- HTML uploads now render as PDF with a widened page so the right edge isn't clipped
- Handle blank pages correctly during line extraction and page separation
- Guard broker publishes against half-open RabbitMQ connections
- On-prem deployments now read the maximum upload file size directly from the
MAX_FILE_SIZEenvironment variable (in MB, default 300) at request time. Previously the configured limit was cached per subscription and not refreshed, so changes toMAX_FILE_SIZEdid not take effect until the cache was cleared. - Helm chart: refreshed the on-prem values schema, and removed the non-functional memory-based autoscaling on the page-extraction-document-insights worker (it was pinned to a single replica and never scaled); the worker now uses default CPU-based autoscaling
Upgrade Notes
- No action required for the page-extraction-document-insights autoscaling change. The previous memory-based HPA never scaled (pinned to a single replica); the worker now uses default CPU-based autoscaling.
v2.60.0 — April 30, 2026
New Features
- Centralized OCR billing proxy: outgoing billing now routes through the Unstract License Portal instead of Azure Cognitive Services, making on-prem OCR cloud-agnostic and removing the need to whitelist Azure endpoints
list-onprem-imagesscript for enumerating required images- Private registry override support for on-prem deployments
Fixes & Improvements
- On-prem Helm chart improvements: removed unused standalone
tesseractandazure-ocrsubcharts (no action needed unless you referenced these subcharts directly in yourvalues.yaml)
Upgrade Notes
azureOcrBilling.endpointandapiKeyare no longer used and can be removed from yourvalues.yaml(leaving them is harmless).- Outbound whitelisting simplified: previously both the Unstract License Portal (
LICENSE_PORTAL_URL) and Azure Cognitive Services endpoints had to be whitelisted. From v2.60.0 onwards, only the Unstract License Portal needs to be whitelisted. - See Outgoing Data (OCR Billing) for details.
v2.59.0 — April 07, 2026
New Features
- Add allow_rotated_text parameter to skip watermark words in OCR
- API key authentication for on-prem internal API
Fixes & Improvements
- Handle TIFF files with unrecognized palette modes
- Detect PDFs with %PDF signature offset from byte 0
- On-prem Helm chart improvements: clean up values and remove redundant config, add GCP ingress annotation placeholders with nil-safe helper, and add global service account override
v2.57.2 — March 12, 2026
New Features
- Configurable RabbitMQ HA mode with quorum queues
- Multi-AZ deployment support with consolidated values overlay for Redis
Fixes & Improvements
- Limit whisper hash lookup to records from last 7 days
- Make MAX_FILE_SIZE configurable and clean up ingestion cache on exceptions
- Tune tesseract sidecar gunicorn workers and threads
- Change Redis PDB from minAvailable to maxUnavailable
- Remove dead Celery config for thread pool workers
- Remove sidecar readiness probes and add liveness timeout
- Global nodeSelector with per-workload override
- Global registry override support for vLLM and Ollama
v2.56.10 — February 24, 2026
New Features
- Auto-generate container image list for on-prem upgrades
- Add open-source fonts for better PDF conversion fidelity
Fixes & Improvements
- Optimize whisperer_record_v2 update queries with created_at filter
- Patch pdfminer PDFStream.decode() TypeError for valid PDFs
- Split page_extraction_low_cost into dedicated worker and update on-prem defaults
- On-prem Helm chart improvements: split image config into registry + name + tag with global defaults, fix NOTES.txt image helpers, resolve Helm v3.9.0 CI fallback, and remove redundant registry overrides from ollama and vllm
v2.56.4 — February 16, 2026
Fixes & Improvements
- Fix corrupted PDF repair for truncated incremental updates and isolate orphaned form field repair in a subprocess to prevent crashes
- Fix Excel text extraction alignment, stripping, and column width
- Reduce low-cost page extraction concurrency to 1:1 with sidecar
- On-prem Helm chart improvements: support existingSecret in RabbitMQ CRD credential resolution, auto-generate RabbitMQ affinity from global.nodeSelector, fix sidecar secretKeyRef for Azure OCR billing, and update schema with ingress, managed credentials, and annotations
v2.56.2 — February 12, 2026
Fixes & Improvements
- Clean up on-prem Helm values and move sensitive credentials (Azure OCR billing, dashboard) to Kubernetes secrets
v2.56.1 — February 10, 2026
New Features
- Add sidecar worker pools for mode-specific OCR processing
- Add Prometheus metrics endpoint for RabbitMQ queue monitoring
- Shared ConfigMaps and Secrets pattern for Helm chart multi-service configuration
- Rename signature mode to document_insights for better API clarity
Fixes & Improvements
- Improve OCR reliability with HTTP retries, exception handling cleanup, and Kubernetes native sidecars for startup ordering
- Fix Excel extraction issues including percentage formats, literal prefixes in number formats, and multi-line text truncation
- Optimize Celery worker graceful shutdown with background heartbeat thread
- Add idempotency check to prevent duplicate page extraction on RabbitMQ redelivery
- On-prem Helm chart improvements: shared secrets, auto-generated API keys, REQUIRED annotations, Zot pull-through cache, and security defaults
- Set HPA target CPU utilization to 80% for Cast AI vertical optimization
- Add fallback migration for document insights mode page count column
- Detect corrupted PDF streams before page extraction to trigger repair
v2.52.0 — January 19, 2026
New Features
- Dashboard now shows the version of the deployed release
- Find by search for whisper hash and file name in dashboard
Fixes & Improvements
- Fixed memory leak in the worker ingestion component
v2.51.1 — January 13, 2026
New Features
- New signature mode introduced
Fixes & Improvements
- Low confidence word association
- Chart optimizations
- Added preStop hook for Azure OCR container deployments
v2.50.6 — December 11, 2025
Fixes & Improvements
- Redis AOF and buffer settings fine-tuning for better reliability and performance
- Fallback to handle Redis replication delays
- Libreoffice optimization
v2.50.4 — December 1, 2025
Fixes & Improvements
- Reduced OCR verbosity to warning from information level
- Addressed Redis and OCR image vulnerabilities
- Chart optimizations
v2.50.3 — November 20, 2025
New Features
- Introduced line-level confidence
Fixes & Improvements
- General optimizations
- Fixed race condition in the document processing pipeline
- Chart optimizations
- Enhanced table mode to handle complex tabular structures
v2.49.3 — October 22, 2025
New Features
- LLMWhisperer dashboard UX changes and sorted filtering issues
- Improved HTML file text extraction
Fixes & Improvements
- Optimized Docker images and addressed high-priority vulnerabilities
- Improved error handling and exception propagation
- Preserved page ordering by introducing empty text in case pages are sent with zero text
- Added decimal format support and prevented type conversion for shared strings in the Excel formatter
- Optimized inefficient PDF files with excessive XObject duplication in the processing pipeline
v2.47.1 — September 19, 2025
New Features
- Added support for XLSB file processing
Fixes & Improvements
- Zot registry pull-through cache for faster and cost-effective container management in on-prem
- Introduced health check probes for workers
- Enhanced error logging
- Fixed Docker image-level vulnerabilities
- Security fixes
- Added automatic detection and repair of orphaned PDF form fields
- On-prem profile tweaking for better calling performance
- Fixed PDF corruption in native text mode
v2.46.3 — August 13, 2025
Fixes & Improvements
- Graceful shutdown for backend and tesseract services
- Introduced HA for the backend service
- License expiration/call home fixes
- Added gunicorn integration for the backend service to enable production-ready deployment with better performance and stability
v2.46.0 — July 18, 2025
New Features
- Switched to the prefork worker model for improved performance and reliability when processing large files
- Dashboard enhancement: Requested page count is now displayed even before the document processing is complete
- Introduced a page limit — each document can now have a maximum of 1,200 pages processed
- Added validation and error handling for password-protected files
- Implemented multiple processing queues based on mode to ensure fair resource allocation across modes
Fixes & Improvements
- Helm chart optimization: Migrated third-party Docker images to a private registry to avoid Docker Hub pull rate limits. Also added PDB for backend workloads
- Improved fault tolerance: Handled internal worker-level
SIGKILLscenarios that previously caused indefinite system hangs
v2.45.5 — July 10, 2025
Fixes & Improvements
- Fix for handling multi-frame TIFF files
- Optimized XLS extraction
v2.45.4 — July 9, 2025
Fixes & Improvements
- Updated Docker base images to use more stable versions
- Improved Excel file formatting during data extraction
- Added a new parameter to ignore vertical text in source documents
- Resolved a segmentation fault issue occurring under high load with Nuitka-compiled binaries
- Reduced cache TTL from 24 hours to 2 hours
- Switched to synchronous API calls for OCR services to improve reliability
- Enhanced logic to detect PDF file types even when MIME type metadata is missing
- Updated OCR image references to use the Unstract container registry
v2.42.2 — June 10, 2025
Fixes & Improvements
- OCR version upgrade for better accuracy