Version 0.18

Released November 5, 2026

Setting up a new pack is now a guided interview that collects the questions and terms that matter to you, then starts testing the pack for you. This version has required upgrade steps, so read Before you upgrade first.

New

Setup and calibration

• You can now set up a new pack through a guided, full-screen interview that collects the questions you care about and the terms your team uses while you add knowledge, then starts calibrating your pack for you.

  How to use it:

  1. Create a new pack and fill in the setup form.
  2. Upload your documents and connect your data sources. You can add more at any point during the interview.
  3. Answer the interview questions. Each round shows the evidence behind it, and you can leave and pick up at the same question later.
  4. Set your pack’s calendar, including fiscal year start, week start and time zone, and mark any questions that must always be answered correctly.
  5. Check the summary on Review, then select Start calibrating.

• While you calibrate, Pack health now shows a Needs you list of tasks waiting on you, alerts, a forecast of how accurate and fast answers will be, and each test question’s speed and cost across test runs.

• You can now run calibration from any AI tool connected to Edisyl over MCP, including checking status, starting test runs, pausing, updating test questions, editing your team’s terms, and undoing decisions.

• During calibration, you can now overrule a test grade you disagree with.

• You can now start a full calibration test run whenever you want, and re-test a single question without changing your pack’s readiness.

• You can now keep a pack fixed with --static when you create, update or import it, so the knowledge and saved answers you brought in don’t change through automatic learning.

Answers and knowledge

• In preview, you can now add business definitions and rules that Edisyl always applies when it answers, like “leave out bot traffic by default,” though it’s off by default, you turn it on through the API, and rules that limit answers to certain audiences aren’t enforced yet.

  How to use it:

  1. Review the definitions and rules Edisyl has suggested for your pack, because turning this on accepts every open suggestion.
  2. Turn it on for the pack with PUT /v1/packs/{id}/concept-layer.

• You can now add documents in other languages, including Chinese, Japanese and Korean, and ask questions about them, with numbers, dates and units read the way each language writes them.

• When Edisyl declines a question because data is missing, you can now fill the gap right away, because it says which source it needs and offers Model this table or Connect data.

Access and installs

• You can now sync users and groups from your identity provider with SCIM 2.0, the standard most identity providers use for this, and manage the connection and its access token at /admin/groups.

• You can now connect Gemini Enterprise to Edisyl over MCP and ask questions from there.

• You can now install the Edisyl command-line tool on Windows with PowerShell.

• You can now generate a Helm, CloudFormation or Terraform install from one setup wizard, with an Istio ingress option and a full configuration reference.

Improved

Calibration and readiness

• A pack now shows Ready only when its answers stay stable across full test runs, and the pack page, home page, MCP and command line all show the same status.
• Test grading reads answers the way people write them, so “$30M” and “30 million” count as the same and fewer correct answers get marked wrong.
• Edisyl re-checks saved answers between full test runs and re-tests them soon after their expected answer or source document changes, so they stay current.
• Questions people ask that Edisyl gets wrong now become test questions automatically.
• Packs that are still calibrating now get a full test run at least once a day, within your pause and spending limits.
• When you paste test questions, Edisyl now tells you which ones are already on the pack.

Answers

• Edisyl declines fewer correct answers, because it now recognizes more ways a number appears in your sources, such as percentages in tables, figures with units, and text that wraps across lines in a PDF.
• More questions get fast answers from ones Edisyl has already checked, because Edisyl now saves more of its correct new answers for reuse.
• Answers now show each cited figure inside the passage it came from, so you can check it at a glance.
• Edisyl now reads how your tables relate from your database’s own definitions, so it joins them correctly with less setup.
• Questions that don’t say which time period they mean now get a follow-up asking for it.
• When you ask over the API, you now get results as data instead of a five-row text summary.

Your team’s terms

• Edisyl now lists conflicting entries in your team’s terms and nicknames and tells you when one word points to several terms, so you can fix them.
• Filtering on values in your data, like a region or a status, now works without adding each value to your terms first.
• Nicknames Edisyl adds on its own now stay closer to wording your team would actually use.

Documents

• Tables in PDFs, including tables without borders and tables on scanned pages, now come through intact, with their headers kept.
• Scanned PDFs are read more accurately, including small scans and numbers in tables, and keep their readable pages even when one table page can’t be read.
• Uploads waiting to be processed now appear in Files and notes and update as they progress.
• Uploading a document no longer changes your pack instructions, which you now write yourself in six sections or let the setup interview write for you.

Speed and reliability

• Edisyl stays responsive under heavy use, because we fixed the causes of occasional freezes and sped up slow lookups.
• Test runs no longer take capacity from questions people are asking.
• Calibration that stalls now restarts on its own and tells you when it was stuck.

Roles and access

• Organization roles are now admin, operator and consumer, replacing owner and member.
• Source managers can now remove a data source from a pack.
• Pack creators can now give roles to many people at once when they create a pack.
• People connecting over MCP now see only the tools they have permission to use.
• We tightened access checks on organization settings and on reading values from your data.

Before you upgrade

Action required

• Every deployment now sends AI model requests through a gateway that uses the OpenAI API format, such as LiteLLM, and the AI providers page shows that gateway instead of your model vendor keys.

  1. Set up an OpenAI-compatible gateway with access to your model providers.
  2. Point Edisyl at the gateway, following the configuration reference.
  3. After upgrading, check that the AI providers page shows your gateway.

• Images now come from the edisyl-ga registry and use release tags like 0.18.0 instead of prd-<sha> tags. The registry and the tag only work together, so overriding one without the other leaves pods unable to pull their images.

  1. If your values file sets an image tag for any app, remove it to use the chart’s defaults, or change it to 0.18.0.
  2. If your values file sets images.registry, remove it or change it to the edisyl-ga registry.

• The sign-in service now checks when it starts that its token issuer matches your public sign-in address, and won’t start if they differ.

  1. If you set GOTRUE_JWT_ISSUER yourself, check that it is your AUTH_PUBLIC_URL followed by /auth/v1. If you don’t set it, there’s nothing to do.

• If you used brainstorm to generate test questions for an existing pack, it now runs only during pack setup.

  1. Add test questions to existing packs directly, or paste them in from your own list.

Good to know

• Calibration can now spend up to $600 a day across your deployment by default, up from $250, and up to $150 a day on any one pack. To keep the old limit, set PACK_AUTO_CALIBRATE_TOTAL_DAILY_USD to 250 under edisyl.api.env in your values file.
• If you want groups from your identity provider to grant roles at sign-in, turn it on by setting IDP_GROUP_MAPPINGS_ENABLED to true under edisyl.edisyl.env. It’s off by default.
• Edisyl now clears out stored data from finished workflow runs after 7 days by archiving it to your object storage, but only reports what it would clear until you turn it on. To turn it on, set WORKFLOW_PAYLOAD_RETENTION_DRY_RUN to false under edisyl.api.env.
• If you override INNGEST_DEV, the API and worker now won’t start with it turned on, and the worker won’t start without both Inngest keys. The chart’s defaults already meet both checks.

Known issues

• An answer can ignore a specific name in your question, like a company, product or condition, and return the total for everything (or 0 when your data spells the name differently) without saying it left the name out.

  Workaround: Check the filters shown with the answer, and add the name to your team’s terms or mark the column as searchable.

• An answer can ignore a time window in your question, so “last 5 business days” returns an all-time figure that looks like the one you asked for.

  Workaround: Use exact dates, like “from Oct 1 to Oct 31.”

• A saved answer can lock a period like “last month” to the date it was first answered.

  Workaround: Ask the question again instead of reusing the saved answer.

• Short codes in large tables may not be recognized, so a question like “Phase 3 trials” can return 0.

  Workaround: Use the value exactly as it’s stored, or add it to your team’s terms.

• Results stop at 50 rows, and you can’t get the rest.

  Workaround: Ask for a summary or a narrower slice of the data.

• Month and period boundaries may use UTC instead of your pack’s time zone, so counts near the start or end of a month can be slightly off.

  Workaround: Set your pack’s time zone during setup, and double-check counts near month boundaries.

• Months that haven’t happened yet can drop out of your pack’s terms, so “storms in November” stops matching until November’s data arrives.

  Workaround: Ask with the month’s dates instead of its name.

• The accuracy figure in Pack health may not include grades you corrected in Review.

  Workaround: Check individual test run results instead of the headline figure.

• Under heavy load, you may see slow responses or a failed step.

  Workaround: Try again.

• A question waiting on your answer to a follow-up fails if your deployment is upgraded in the meantime, and we’ve fixed this for 0.19.

  Workaround: Ask the question again after the upgrade.

Get this version

Version 0.18 is no longer the latest. To install or upgrade, use version 0.18.2.