← Blog

· Hive AI

Knolo CLI 6.2.9: knolo ask Answers Only from Verified Hub Excerpts

Knolo CLI 6.2.9 adds knolo ask: cited Hub excerpts, SHA-256 checked, sent only to the model endpoint you choose, with no model call when evidence is missing.

Knolo is the open-source, local-first Knowledge Image runtime and format: Apache-2.0, published on npm as @knolo/core. On 17 November 2026 the Knolo CLI moved to @knolo/cli 6.2.9, and that release adds one new command: knolo ask. It takes a question, pulls cited excerpts from Knolo Hub, drops any excerpt whose SHA-256 does not match, and sends only what is left to a model endpoint you choose.

This post covers what knolo ask does, the flags it requires, and the cases where it deliberately refuses to call a model.

What Knolo CLI 6.2.9 adds

The work landed in the Knolo core repository on GitHub as two merged pull requests and a version stamp:

  • #198, ground knolo ask in Hub evidence. The command queries Hub for cited excerpts, checks each excerpt hash, and sends the rest to an OpenAI-compatible chat endpoint.
  • #199, choose categories for knolo ask. The command now requires --category or --all, and matched skills are checked and printed before knowledge hits.
  • #201, stamp 6.2.9. The repository records that npm latest for @knolo/cli is 6.2.9, and that 6.2.8 has no ask.mjs and no --category.

@knolo/cli 6.2.9 depends on @knolo/core ^6.2.3 and @knolo/exchange ^0.1.1. Nothing else moved on 17 November: @knolo/core stays at 6.2.3.

Install the Knolo CLI

npm install -g @knolo/cli@6.2.9
knolo ask --help

If you are new to the CLI, start with the Knolo installation guide and the Knolo CLI reference. The registry side is described in the Knolo Hub docs.

Knolo ask, step by step

knolo ask "what is the refund window?" --category business --model llama3.2
knolo ask "how does rust ownership work" --category coding --category math --evidence
knolo ask "what changed overnight" --all --model llama3.2
  1. Pick a scope. Pass --category one or more times, or pass --all to search the entire Hub. The categories are coding, math, science, news, business, security, writing and general. With neither flag, or with both, the command exits before any network call.
  2. Query Hub. The question goes to POST /api/v1/knowledge/query on the registry, which is https://hub.knolo.dev unless you pass --hub/--registry or set KNOLO_HUB_URL. --limit accepts 1 to 20.
  3. Check every excerpt. Each knowledge hit carries publisher, slug, version, objectId, the excerpt text and an excerptSha256. The CLI recomputes SHA-256 over the excerpt and drops any hit that does not match. Skills go through the same check and must also carry a skillSha256.
  4. Call your model. The surviving skills and excerpts go to <base-url>/chat/completions with a fixed system prompt: answer only from the excerpts and skill instructions, cite each claim as publisher/slug@version plus the object id, and use no outside knowledge.
  5. Print the answer and the citations for every skill and excerpt that was sent.

Where Knolo ask sends your question and your key

The default model endpoint is a local Ollama server at http://127.0.0.1:11434/v1. Set KNOLO_MODEL_BASE_URL and KNOLO_MODEL (or --base-url and --model) to use another local server or a cloud gateway. A base URL that includes a username or password is rejected.

KNOLO_MODEL_API_KEY is sent only to that model endpoint. Hub receives the question, but it never receives the model key or the answer, and knolo ask does not install a matched skill. If you are logged in to the same registry with knolo login, the stored Hub token is sent to Hub. Otherwise the query goes out without one.

When Knolo ask does not call a model

The command fails closed rather than letting a model improvise:

  • No --category and no --all: no network call at all; the command exits with an error.
  • Hub request fails: no model call; exit code 1.
  • No excerpt or skill survives the hash check: no model call; exit code 1.
  • --evidence: no model call; skills, then excerpts, are printed with their hashes; exit code 0.
  • Model endpoint down or returns an error: the call is tried, then the verified excerpts are printed instead; exit code 2.
  • Model answers: the answer and its citations are printed; exit code 0.

If only a skill survives the check, the model still runs with that skill text. --evidence is the quickest way to see exactly what a model would be given.

What the Knolo excerpt check proves, and what it doesn't

The SHA-256 check proves that each excerpt arrived byte-for-byte as Hub described it. It does not prove the excerpt against a pack's state root, and it does not make the model's answer true. When you need that stronger guarantee, install the pack by digest with knolo add, which verifies the manifest, the Blob bytes, the digest, the size and the local image structure, then query it locally and keep a receipt. The Knolo verification docs and the Knolo query receipts docs cover that path.

knolo ask is the quick, cited answer. A pinned Knowledge Image is the record you can re-check later.

Learn more about Knolo

Knolo is built by Hive AI and released as open source under Apache-2.0.