Knowledge base

Les descriptions des endpoints et des champs restent en anglais, telles que l'API les publie. C'est un choix délibéré : vous lisez ici ce que vous verrez aussi dans les réponses.

GET /v1/kb/collections#

List collections. Every collection on your account, with its document and chunk counts.

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/collections" \
  -H "Authorization: Bearer VL_API_KEY"

Exemple de réponse

{
  "items": [
    {
      "doc_count": 12,
      "id": "col_9f2a1c7d",
      "name": "Τιμοκατάλογος",
      "vector_count": 143
    }
  ]
}

POST /v1/kb/collections#

Create a collection. Creates a collection, a named set of documents your agents can answer from. Collections belong to your account, not to one agent: any number of agents may attach the same collection, and attaching never changes it. Supply id to make the call repeatable (a second create with the same id returns 409), or omit it for a generated one.

Corps de la requête

application/json Schéma : KbCollectionRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/collections" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "default_lang": "el",
  "description": "Price list and billing rules.",
  "name": "Τιμοκατάλογος"
}'

Une réponse réussie renvoie un KnowledgeCollection.

Exemple de réponse

{
  "default_lang": "el",
  "doc_count": 0,
  "id": "col_9f2a1c7d",
  "name": "Τιμοκατάλογος",
  "vector_count": 0
}

DELETE /v1/kb/collections/{id}#

Delete a collection. Deletes an empty collection. A collection that still holds documents returns 409, pass ?cascade=true to delete its documents and remove them from the search index in the same call.

Paramètres

Réponses

Exemple de requête

curl -X DELETE "https://api.voiceland.ai/v1/kb/collections/{id}" \
  -H "Authorization: Bearer VL_API_KEY"

GET /v1/kb/collections/{id}#

Get a collection. Returns one collection by id.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/collections/{id}" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeCollection.

Exemple de réponse

{
  "doc_count": 12,
  "id": "col_9f2a1c7d",
  "name": "Τιμοκατάλογος"
}

PATCH /v1/kb/collections/{id}#

Update a collection. Changes the name, description, or default language. Omitted fields are left alone; the documents inside are untouched.

Paramètres

Corps de la requête

application/json Schéma : KbCollectionPatchRequest

Réponses

Exemple de requête

curl -X PATCH "https://api.voiceland.ai/v1/kb/collections/{id}" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "description": "Τιμές και εκπτώσεις 2026"
}'

Une réponse réussie renvoie un KnowledgeCollection.

Exemple de réponse

{
  "description": "Τιμές και εκπτώσεις 2026",
  "id": "col_9f2a1c7d"
}

GET /v1/kb/collections/{id}/documents#

List documents. One page of the documents in a collection, in id order. The response omits each document's body, fetch a single document for its text. **Paging.** Pass the next_cursor from the previous response back as ?cursor= to get the page after it. **An empty next_cursor is the only signal that you have seen everything**, a short page, including an empty one, is normal and means "ask again": status is applied after each read, so a page can legitimately match nothing and still have more behind it. Follow the cursor until it comes back empty and your account's whole collection is covered, however large it is. Ordering is by document id within the collection, not by when a document was last edited: a page cannot sort documents it has not read, and a listing that claimed an order it could not keep across pages would repeat and skip documents.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/collections/{id}/documents" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbDocumentsResponse.

Exemple de réponse

{
  "items": [
    {
      "chunk_count": 7,
      "id": "doc_4b1e77a0",
      "status": "published",
      "title": "Τιμολόγηση 2026",
      "version": "1.1.0"
    }
  ],
  "next_cursor": "Y29sXzlmMmExYzdkI2RvY180YjFlNzdhMA"
}

POST /v1/kb/collections/{id}/documents#

Add a document. Adds a document to the collection. Five shapes are accepted on this one route: - **inline text**, JSON with title + body. Indexed synchronously; responds **201** with the document. - **file upload**, multipart/form-data with a file part (PDF, DOCX, HTML, Markdown, TXT, CSV) plus optional title, lang, tags, categories, audience, source_authority fields. Send the fields BEFORE the file. - **a web page**, JSON {"url": "https://example.com"}. The page is fetched, extracted and indexed. - **your website**, JSON {"website": "https://example.com"}, and the one to reach for. You do not have to know where your sitemap lives: we read your robots.txt, then try the conventional paths, and crawl the sitemap we find. A site that publishes none is crawled by FOLLOWING ITS LINKS instead, same result, one document made of many pages. A deep link is reduced to the site it belongs to, so pasting the page you had open works. The **202** carries a discovery block saying which sitemap was found and where it was declared, or, when there was none, which addresses were tried and that links will be followed. Takes the same optional scope as a sitemap. - **a sitemap**, JSON {"sitemap": "https://example.com/sitemap.xml"} with an optional scope (allow/deny path prefixes, max_pages). **Leave max_pages out and you get every page**, there is no default cap; a limit applies when you state one or when your plan sets one, and the job says when it stopped short and why. The whole crawl becomes ONE document made of many pages, so each answer still links to the exact page it came from. Add scope.recrawl_interval ("24h", between 1h and 8760h, one hour to one year) to have the source re-fetched on a timer, the re-fetch is conditional, so a page that has not changed costs you nothing, and the schedule only runs while the document is published. It applies to a single url as well as to a sitemap. You send it as a duration string; the document reports it back as a number of nanoseconds. Everything except inline text responds **202** with the draft document AND an ingestion job, poll GET /kb/jobs/{jobID} for progress. Every shape starts as a draft at version 1.0.0 and is NOT searchable until you publish it. **Standardised taxonomy.** Beside the free-form tags and categories, every shape accepts two typed fields that search and the document listing can narrow by: - audience, who the document is written for (["customers", "partners"], «συνεργάτες»). Your own vocabulary, in any language, but it is stored lower-cased with runs of spaces collapsed and matched EXACTLY, so pick the words once and reuse them. At most 8 values, 64 characters each. Nothing is enforced; if you have no opinion of your own, these are the words we suggest so that two collections agree: customers, prospects, partners, employees, agents, public, internal. - source_authority, how much weight the document carries: official (your own published position), verified (third-party material somebody checked), community (written by users), unverified (explicitly not checked). Nothing else is accepted. **Leaving either one out is not a value.** A document with no source_authority has made no claim, and unverified says the opposite, that somebody looked. Filters honour the difference: narrowing a search to unverified returns the documents marked so, never the ones nobody has labelled. Nothing on the platform fills these in for you. If you have been keeping the same information in metadata (metadata.audience and the like), it stays exactly where it is and keeps working, it is not read as a fallback and nothing was migrated, because a value nobody validated should not silently become a retrieval filter. Move it across when you are ready, one document at a time.

Paramètres

Corps de la requête

application/json Schéma : KbDocumentRequest

Corps de la requête

multipart/form-data

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/collections/{id}/documents" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "body": "# Τιμολόγηση\n\nΗ χρέωση γίνεται ανά λεπτό ομιλίας…",
  "categories": [
    "Τιμολόγηση/Εκπτώσεις"
  ],
  "lang": "el",
  "tags": [
    "pricing"
  ],
  "title": "Τιμολόγηση 2026"
}'

Une réponse réussie renvoie un KnowledgeDocument.

Exemple de réponse

{
  "collection_id": "col_9f2a1c7d",
  "id": "doc_4b1e77a0",
  "status": "draft",
  "version": "1.0.0"
}

DELETE /v1/kb/documents/{docID}#

Delete a document. Deletes the document and everything derived from it, in order: any ingestion still running is cancelled, its passages are removed from the search index, the stored copy of your file or crawled pages is erased, and then the document itself. Permanent, archive instead if you may want the text back.

Paramètres

Réponses

Exemple de requête

curl -X DELETE "https://api.voiceland.ai/v1/kb/documents/{docID}" \
  -H "Authorization: Bearer VL_API_KEY"

GET /v1/kb/documents/{docID}#

Get a document. Returns one document including its full text. A document with a re-crawl schedule also carries recrawl_state: last_crawled_at and next_due_at. Read the ABSENCE of next_due_at beside a non-zero recrawl_interval as meaningful rather than as an error, only a published document is ever re-fetched, so that combination is the usual answer to "I set a schedule and nothing is happening". The listing does not carry it; ask for the document.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/documents/{docID}" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeDocument.

Exemple de réponse

{
  "id": "doc_4b1e77a0",
  "status": "published",
  "title": "Τιμολόγηση 2026",
  "version": "1.1.0"
}

PATCH /v1/kb/documents/{docID}#

Update a document. Edits the title, text, language, tags, categories, audience, source authority, metadata, validity window or re-crawl schedule. **Sending body is a redaction, and it is destructive.** The text you send becomes the whole content of the document, and everything derived from the old text is destroyed rather than left to reappear: - the passages already in search are removed, so the old text stops being answerable the moment you save; - a published document drops back to draft, publish again to make the new text live. A draft stays a draft, and editing anything other than body leaves a published document published; - **on an uploaded or crawled document, the stored copy of your file or crawled pages is erased, any ingestion still running on it is cancelled, and the document's source becomes inline.** From then on the text on the row is the only content of record: re-index and publish rebuild from what you sent, never from the file you originally uploaded or the page we crawled. source.url is kept as provenance for citations, but nothing re-fetches it. That last point is the difference between an edit and a redaction, and it is one-way. If you need the original bytes back, upload them again as a new document. If you only meant to correct a typo in a file you still want us to hold, edit the file and re-upload it instead. **Moving the validity window also takes a published document out of search**, and for the same reason a text edit does: the window is stamped into every passage when you publish, so a row whose dates moved while its passages still carry the old ones is a document that answers on a day it should not. It is a NARROWER action than a body edit, nothing stored is destroyed and the source is not flipped to inline, the document simply drops back to draft, and publishing again re-stamps the passages. Send "" on either side to clear it. recrawl_interval sets or clears the schedule (a Go duration, "24h", between 1h and 8760h, one hour to one year; "" stops it). It applies to documents whose source is a url or a sitemap, and only a published one is ever re-fetched. audience and source_authority are the standardised taxonomy, see the create call for what the values mean. Send "audience": [] or "source_authority": "" to take a claim back off a document; there is no word in the vocabulary for "we have not said", so clearing is how you say it. **Re-labelling does NOT take a published document out of search**, it is a label, not content, and the same is true of tags and categories. The passages already in search keep the labels they were published with until you publish again, so a search narrowed by the new label finds the document only after the next publish. metadata replaces the map you own. Keys the platform manages, anything starting crawl_ or vl_, plus filename and content_type, cannot be set and are preserved across the replacement (400 reserved_metadata_key if you send one). The same rule applies when you CREATE a document. metadata.audience and friends are NOT reserved and are NOT read as a fallback for the typed fields: if you have been keeping this information there it keeps working untouched.

Paramètres

Corps de la requête

application/json Schéma : KbDocumentPatchRequest

Réponses

Exemple de requête

curl -X PATCH "https://api.voiceland.ai/v1/kb/documents/{docID}" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "body": "# Τιμολόγηση\n\nΑναθεωρημένες τιμές…"
}'

Une réponse réussie renvoie un KnowledgeDocument.

Exemple de réponse

{
  "id": "doc_4b1e77a0",
  "source": {
    "kind": "inline"
  },
  "status": "draft",
  "version": "1.1.0"
}

POST /v1/kb/documents/{docID}/archive#

Archive a document. Takes the document out of search while keeping the text and its history. Safe to repeat. Publish again to bring it back. The passages are removed from the index before the document is marked archived, and the removal is attempted whether or not the document is recorded as indexed, so an archive can only ever end with less in search than it started with, never more.

Paramètres

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/archive" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeDocument.

Exemple de réponse

{
  "chunk_count": 0,
  "id": "doc_4b1e77a0",
  "status": "archived"
}

GET /v1/kb/documents/{docID}/chunks#

Preview passages. Shows how the document splits into retrievable passages, with the heading each one sits under. count is what the current text would produce; indexed_count is what search actually holds, they differ when the text was edited after the last publish.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/documents/{docID}/chunks" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbChunksResponse.

Exemple de réponse

{
  "count": 7,
  "document_id": "doc_4b1e77a0",
  "indexed_count": 7,
  "items": [
    {
      "chunk_id": "doc_4b1e77a0#0",
      "index": 0,
      "section": "Τιμολόγηση",
      "text": "Η χρέωση γίνεται ανά λεπτό ομιλίας…"
    }
  ],
  "status": "published",
  "version": "1.1.0"
}

POST /v1/kb/documents/{docID}/publish#

Publish a document. Makes the document answerable: the text is split into passages, indexed for search, and the version's minor number is bumped so an answer can name the exact revision it used. Safe to repeat, republishing replaces the previous passages rather than duplicating them. Publishable from any state: a draft, a document **in review** (this is the approval step), a published one you are re-indexing, or an archived one you are bringing back. An inline document is indexed inside the call (**200**). An uploaded or crawled one cannot be, its content is not on the row, so publishing it responds **202** with {document, job} and the passages appear once the job reports indexed.

Paramètres

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/publish" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeDocument.

Exemple de réponse

{
  "chunk_count": 7,
  "id": "doc_4b1e77a0",
  "status": "published",
  "version": "1.1.0"
}

POST /v1/kb/documents/{docID}/recrawl#

Fetch this document's source again. Goes back to the address this document came from and reads it again, because the PAGE changed, not because we did. **This is the other half of re-index, and the difference matters.** /reindex rebuilds passages from the copy we already keep, so it is right when our chunking or embedding model improved and useless when your website changed. /recrawl is the opposite: it re-fetches, and the fetch is conditional, a page that has not changed costs you nothing. The kind of walk is the document's own: a document built from a sitemap re-walks that sitemap, one built by following links walks them again, and a single page re-fetches that page. You do not choose, and cannot: it is what the document is made of. Refused with **400** not_a_remote_document when there is no address to go back to, text you typed, a file you uploaded, a bucket import. One job at a time per document, because this one reaches out to somebody's web server.

Paramètres

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/recrawl" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbIngestResponse.

Exemple de réponse

{
  "document": {
    "id": "doc_4b1e77a0",
    "status": "published"
  },
  "job": {
    "id": "job_1c4f9ab2",
    "kind": "sitemap",
    "status": "queued"
  }
}

POST /v1/kb/documents/{docID}/reindex#

Re-index a document. Queues the document to be extracted, split and embedded again, after the platform's chunking improves, or once you have corrected the collection's language. It re-reads the copy we already keep of your file or crawled pages, so re-indexing never hits your website again, which is exactly why it will NOT see a page you have edited. Use /recrawl for that.

Paramètres

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/reindex" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbIngestResponse.

Exemple de réponse

{
  "document": {
    "id": "doc_4b1e77a0",
    "status": "published"
  },
  "job": {
    "id": "job_1c4f9ab2",
    "kind": "reindex",
    "status": "queued"
  }
}

POST /v1/kb/documents/{docID}/review-flag/dismiss#

Dismiss a review flag. Closes a review flag by saying the document was not at fault. **A reason is required**: the flag exists because a human confirmed an answer was wrong, so a dismissal is a claim that the cause was somewhere else, and one with no reason attached is indistinguishable from clearing the queue. The other way out is to publish the document again, which closes the flag as resolved with no extra call: republishing states its case by changing the content. A decision is one-shot. What reopens a closed flag is a NEW confirmed-wrong report naming the same document.

Paramètres

Corps de la requête

application/json Schéma : KbReviewDismissRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/review-flag/dismiss" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "dismissed_by": "maria@acme.example",
  "reason": "The document is correct; the agent quoted the wrong section."
}'

Une réponse réussie renvoie un KnowledgeReviewFlag.

Exemple de réponse

{
  "collection_id": "col_9f2a1c7d",
  "document_id": "doc_4b1e77a0",
  "report_count": 3,
  "resolution_note": "The document is correct; the agent quoted the wrong section.",
  "resolved_by": "maria@acme.example",
  "state": "dismissed"
}

POST /v1/kb/documents/{docID}/submit#

Submit a document for review. Marks a draft as ready for somebody to check before it goes live: the status becomes in_review. Nothing about search changes, a document in review is no more answerable than a draft, so this is a hand-off, not a publish. Only a **draft** can be submitted. Repeating the call on a document already in review succeeds and changes nothing. A **published** document cannot be sent back for review by this call: its passages are live in search, and a status change alone would leave them answering. Edit its text, that takes it out of search and returns it to draft, then submit. Archiving is not a route back: an archived document cannot be submitted either, and the only way out of archived is to publish it.

Paramètres

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/submit" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeDocument.

Exemple de réponse

{
  "id": "doc_4b1e77a0",
  "status": "in_review",
  "version": "1.0.0"
}

GET /v1/kb/documents/{docID}/variants#

List a document's languages. Every language the document holds, the primary one first. The primary is in the list because search falls back to it, a language picker built from this list offers exactly the choices search can honour.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/documents/{docID}/variants" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbVariantsResponse.

Exemple de réponse

{
  "document_id": "doc_4b1e77a0",
  "items": [
    {
      "chunk_count": 7,
      "lang": "el",
      "primary": true
    },
    {
      "chunk_count": 6,
      "lang": "en",
      "primary": false
    }
  ],
  "primary_lang": "el"
}

POST /v1/kb/documents/{docID}/variants#

Add or replace a language version. Stores another language of the same document. A variant carries content only, the status, version, tags, categories and validity window stay on the document, so there is one thing to review and publish rather than one per language. Search prefers the language the person is searching in and falls back to the document's own, so a document with no Greek version still answers a Greek question, in the language it has. Adding a language to a **published** document takes it out of search until you publish again, exactly as editing its text does: what search holds is stamped when you publish, and quietly re-indexing inside this call would spend an embedding batch you did not ask for. Available for documents with inline text. A document ingested from a file, a page or a sitemap gets its language from its source.

Paramètres

Corps de la requête

application/json Schéma : KbVariantRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/variants" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "body": "# Pricing\n\nBilling is per minute of speech…",
  "lang": "en",
  "title": "Pricing 2026"
}'

Une réponse réussie renvoie un KbVariantResponse.

DELETE /v1/kb/documents/{docID}/variants/{lang}#

Delete one language. Removes that language and its passages from search. The other languages are untouched, the removal names this language only. The primary language cannot be deleted this way; deleting the document removes it.

Paramètres

Réponses

Exemple de requête

curl -X DELETE "https://api.voiceland.ai/v1/kb/documents/{docID}/variants/{lang}" \
  -H "Authorization: Bearer VL_API_KEY"

GET /v1/kb/documents/{docID}/variants/{lang}#

Get one language. Returns that language's title and text, or 404 if the document does not have it. It does NOT fall back to the primary, a console asking "is there a Greek version" has to be able to get a no.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/documents/{docID}/variants/{lang}" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbVariantResponse.

Exemple de réponse

{
  "body": "# Pricing…",
  "chunk_count": 6,
  "lang": "en",
  "primary": false,
  "title": "Pricing 2026"
}

PATCH /v1/kb/documents/{docID}/variants/{lang}#

Update one language. Changes that language's title or text. The language itself cannot be changed, delete it and add the other one. Editing the primary language goes through PATCH /kb/documents/{docID} instead, which is the call that also removes the old text from search and from storage.

Paramètres

Corps de la requête

application/json Schéma : KbVariantPatchRequest

Réponses

Exemple de requête

curl -X PATCH "https://api.voiceland.ai/v1/kb/documents/{docID}/variants/{lang}" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "body": "# Pricing\n\nBilling is per minute…"
}'

Une réponse réussie renvoie un KbVariantResponse.

Exemple de réponse

{
  "chunk_count": 0,
  "lang": "en",
  "primary": false
}

POST /v1/kb/documents/{docID}/withdraw#

Take a document out of review. Returns a document in review to draft, the same call whether the author is withdrawing it or a reviewer is sending it back. Nothing about search changes. Only a document **in review** can be withdrawn. Repeating the call on a draft succeeds and changes nothing. A published document is not withdrawn but archived.

Paramètres

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/documents/{docID}/withdraw" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeDocument.

Exemple de réponse

{
  "id": "doc_4b1e77a0",
  "status": "draft",
  "version": "1.0.0"
}

POST /v1/kb/import-inline#

Import an agent's inline knowledge. Turns the free-text knowledge you gave an agent into a knowledge-base document, so it can be searched, versioned, cited and shared with other agents instead of being re-sent to the model on every single turn. **Your agent is not changed.** The inline text stays exactly where it is and keeps working, so you can import, publish, attach, listen to the agent answer from search, and only then remove the inline copy by hand. **Safe to repeat.** The document id is derived from the agent name, so calling it again updates the same document, and if the text has not changed, nothing at all is written (changed: false). Set attach: true to also add the collection to the agent's knowledge sources. audience and source_authority set the standardised taxonomy (see the create call). Changing either one counts as a change, so a re-import that only re-labels the document is written, and, since it changes a label rather than the text, it does not take a published document out of search.

Corps de la requête

application/json Schéma : KbImportInlineRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/import-inline" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "agent": "front-desk",
  "attach": true,
  "collection_id": "col_9f2a1c7d"
}'

Une réponse réussie renvoie un KbImportInlineResponse.

Exemple de réponse

{
  "agent": "front-desk",
  "attached": true,
  "changed": true,
  "created": true,
  "document": {
    "id": "inline_front-desk",
    "status": "draft",
    "version": "1.0.0"
  },
  "source": "system_prompt"
}

GET /v1/kb/jobs#

List ingestion jobs. One page of your ingestion jobs, what is queued, running, indexed or failed. A failed job carries a one-line reason. Jobs are kept for 30 days; the document row is the lasting record. **Paging.** Pass the next_cursor from the previous response back as ?cursor= to get the page after it. **An empty next_cursor is the only signal that you have seen everything**, a short page, including an empty one, is normal and means "ask again": status is applied after each read, so a page can legitimately match nothing and still have more behind it. On this listing that is the ordinary case rather than the edge one, because asking for the failed jobs filters a history that is mostly queued ones. Follow the cursor until it comes back empty and your whole job history is covered, however much you have ingested. Ordering is by key, collection, then document, then job id, not by when a job was created: a page cannot sort jobs it has not read, and a listing that claimed an order it could not keep across pages would repeat and skip jobs. Sort a page yourself for an activity feed, or narrow to one document_id, whose whole history is a handful of rows.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/jobs" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbJobsResponse.

Exemple de réponse

{
  "items": [
    {
      "attempts": 1,
      "document_id": "doc_4b1e77a0",
      "id": "job_1c4f9ab2",
      "kind": "sitemap",
      "progress": {
        "chunks_written": 96,
        "pages_done": 17,
        "pages_found": 42
      },
      "status": "processing"
    }
  ],
  "next_cursor": "Y29sXzlmMmExYzdkI2RvY180YjFlNzdhMCNqb2JfMWM0ZjlhYjI"
}

DELETE /v1/kb/jobs/{jobID}#

Cancel an ingestion job. Asks a queued or running job to stop. Best-effort: a job already writing its last passages may still finish, so the response returns the job so you can see which happened. A job that has already finished returns 409.

Paramètres

Réponses

Exemple de requête

curl -X DELETE "https://api.voiceland.ai/v1/kb/jobs/{jobID}" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeJob.

Exemple de réponse

{
  "id": "job_1c4f9ab2",
  "status": "cancelled"
}

GET /v1/kb/jobs/{jobID}#

Get an ingestion job. One job with its live progress counters. Pass collection_id and document_id together when you have them (the 202 that created the job returned both) to make this a direct lookup instead of a walk.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/jobs/{jobID}" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KnowledgeJob.

Exemple de réponse

{
  "document_id": "doc_4b1e77a0",
  "id": "job_1c4f9ab2",
  "kind": "upload",
  "progress": {
    "chunks_written": 12,
    "pages_done": 1,
    "pages_found": 1
  },
  "status": "indexed"
}

GET /v1/kb/object-sources#

List your object-storage sources. Every bucket you have registered. Secrets are never included, has_secret says one is stored, nothing returns its value.

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/object-sources" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbObjectSourcesResponse.

Exemple de réponse

{
  "items": [
    {
      "access_key_id": "AKIA…",
      "bucket": "acme-docs",
      "endpoint": "s3.eu-central-1.amazonaws.com",
      "has_secret": true,
      "id": "objsrc_a1b2c3d4",
      "prefix": "handbook/"
    }
  ]
}

POST /v1/kb/object-sources#

Register your own object storage. Registers a bucket of yours, AWS S3 or anything that speaks the S3 API, so you can import documents straight out of it instead of uploading them one at a time. **`secret_access_key` is write-only.** It is stored encrypted and is never returned by this call, by the list, by the GET, or by the secrets API, there is no request that reads it back. `has_secret` is how you confirm one is stored. To change it, PATCH this source with a new `secret_access_key`; that is the only way to rotate it. **`endpoint` is a host, not a URL** (`s3.eu-central-1.amazonaws.com`, `objects.example.gr`). We always connect over https on port 443, because anything else would put your access key on the wire in the clear. An `https://example.com prefix is accepted and stripped; a path, a port we do not reach, credentials in the endpoint, or an address on a private network are refused (`endpoint_not_allowed`), and the same check runs again, against the address we actually connect to, every time an import runs. **Give us the narrowest credentials that work.** A read-only key scoped to `prefix` is all an import needs: we only ever LIST and GET. `prefix` is also a boundary, an import may narrow it, never widen it, so a source registered at `handbook/` can never be made to read the rest of the bucket.

Corps de la requête

application/json Schéma : KbObjectSourceRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/object-sources" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "access_key_id": "AKIA…",
  "bucket": "acme-docs",
  "endpoint": "s3.eu-central-1.amazonaws.com",
  "name": "Handbook bucket",
  "prefix": "handbook/",
  "region": "eu-central-1",
  "secret_access_key": "…"
}'

Une réponse réussie renvoie un KbObjectSourceView.

Exemple de réponse

{
  "access_key_id": "AKIA…",
  "bucket": "acme-docs",
  "endpoint": "s3.eu-central-1.amazonaws.com",
  "has_secret": true,
  "id": "objsrc_a1b2c3d4",
  "name": "Handbook bucket",
  "prefix": "handbook/",
  "region": "eu-central-1"
}

DELETE /v1/kb/object-sources/{sourceID}#

Remove an object-storage source. Removes the registration and erases the stored credential. **It refuses while documents imported from this source still exist** (409 object_source_in_use, naming one of them). That is deliberate: those documents are your indexed content, and deleting a piece of configuration must not destroy them. Delete the documents you no longer want first, deleting a document removes its passages and the copies of your files we kept, and then remove the source. Nothing in your own bucket is touched, ever. We only ever read from it.

Paramètres

Réponses

Exemple de requête

curl -X DELETE "https://api.voiceland.ai/v1/kb/object-sources/{sourceID}" \
  -H "Authorization: Bearer VL_API_KEY"

GET /v1/kb/object-sources/{sourceID}#

Get one object-storage source. The registration as stored. has_secret is checked against the stored credential on this call, so it is the honest answer to "is this source usable"; the secret itself is never returned.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/object-sources/{sourceID}" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbObjectSourceView.

Exemple de réponse

{
  "bucket": "acme-docs",
  "has_secret": true,
  "id": "objsrc_a1b2c3d4"
}

PATCH /v1/kb/object-sources/{sourceID}#

Update an object-storage source. Changes what is registered. Omitted fields are left alone. **This is how you rotate the credential**: send a new secret_access_key and it replaces the stored one. There is no way to clear it, a source with no credential cannot be imported from, and no way to read the old one back. Documents already imported from this source keep their content. The next import (or re-index that has to re-read the bucket) uses the new settings, so changing the endpoint or bucket redirects future reads without touching what is already indexed.

Paramètres

Corps de la requête

application/json Schéma : KbObjectSourcePatchRequest

Réponses

Exemple de requête

curl -X PATCH "https://api.voiceland.ai/v1/kb/object-sources/{sourceID}" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "secret_access_key": "…"
}'

Une réponse réussie renvoie un KbObjectSourceView.

Exemple de réponse

{
  "bucket": "acme-docs",
  "has_secret": true,
  "id": "objsrc_a1b2c3d4"
}

POST /v1/kb/object-sources/{sourceID}/import#

Import from your object storage. Reads every object under the prefix and indexes them, through exactly the same extraction and chunking an upload goes through, PDFs, Word files, Markdown, HTML, CSV and plain text. **One import is one document**, made of every file under the prefix, the same way a sitemap crawl produces one document made of many pages. Each passage still carries the file it came from (s3://bucket/key) so an answer cites the exact file. If you want one document per file, import a narrower prefix per document. Responds **202** with {document, job}: the document exists immediately as a draft, and the files become searchable once the job reports indexed and you publish. Poll GET /kb/jobs/{id}. **Re-importing is cheap and safe, and it is how you pick up changes.** Call this again with the id of a document this same source already produced and it REFRESHES that document rather than making a new one: every file whose ETag has not changed is served from the copy we kept, not downloaded from your bucket, and not re-embedded, and when nothing under the prefix has changed at all the job finishes having written nothing. A refresh is not an edit: the title, the labels and the published status you set are left alone, and the prefix stays whatever that document was imported at (the request's prefix is ignored). Poll or schedule this as often as you like. POST /kb/documents/{id}/reindex is a different thing, it rebuilds passages from the copies we already hold and never looks at your bucket, so it will not see a new or edited file. Your plan's limits are the upload limits: the same per-file size cap, and a per-import ceiling on how many files one import reads. Hitting the ceiling is reported as a warning on the job rather than a failure, narrow the prefix and import the rest separately. audience and source_authority are the standardised taxonomy and mean exactly what they mean on the create call, a bucket of files is one document, so the labels are the document's, not the files'.

Paramètres

Corps de la requête

application/json Schéma : KbObjectImportRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/object-sources/{sourceID}/import" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "collection_id": "col_9f2a1c7d",
  "prefix": "handbook/policies/",
  "tags": [
    "hr"
  ],
  "title": "Employee handbook"
}'

Une réponse réussie renvoie un KbIngestResponse.

Exemple de réponse

{
  "document": {
    "id": "doc_4b1e77a0",
    "source": {
      "kind": "object",
      "object_prefix": "handbook/policies/",
      "object_source_id": "objsrc_a1b2c3d4",
      "url": "s3://acme-docs/handbook/policies/"
    },
    "status": "draft"
  },
  "job": {
    "id": "job_7c1d",
    "kind": "object",
    "status": "queued"
  }
}

GET /v1/kb/review-flags#

List documents flagged for review. One page of the review queue: documents a human traced a wrong answer back to. A flag is raised when somebody resolves an inaccuracy report as confirmed_wrong and names the document that caused it (POST /feedback/reports/{id}/resolve), and it carries the ids of the reports behind it so an author can read why the document is suspect. **Ordering is by how many reports point at each document**, most first, then by collection and document id. That is a ranking rather than a key order, so the read is bounded: truncated: true means the ranking is over the flags examined and not over every flag on your account. Use collection_id to narrow it, or the analytics route for the whole picture. **Paging.** Pass the next_cursor from the previous response back as ?cursor=. **An empty next_cursor is the only signal that you have seen everything**: state is applied after each read, so a short page, including an empty one, is normal and means "ask again". A flag leaves the queue in one of two ways: publish the document again (the fix speaks for itself, and the flag closes as resolved), or dismiss it with a reason. A new confirmed-wrong report on a closed flag reopens it.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/review-flags" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbReviewFlagsResponse.

Exemple de réponse

{
  "items": [
    {
      "agent_name": "front-desk",
      "collection_id": "col_9f2a1c7d",
      "document_id": "doc_4b1e77a0",
      "reason": "confirmed_wrong_answer",
      "report_count": 3,
      "report_ids": [
        "fbr_9d8c7b6a",
        "fbr_5e4d3c2b",
        "fbr_1a2b3c4d"
      ],
      "state": "open"
    }
  ],
  "next_cursor": "",
  "truncated": false
}

GET /v1/kb/review-flags/analytics#

Which documents cause wrong answers. Ranks your documents by how many answers a human confirmed wrong and traced back to them. This is the number that says where your knowledge is actually wrong, as opposed to where the agent is. knowledge_fault_rate is the share of confirmed-wrong answers a reviewer traced to a document at all; linked_to_document beside confirmed_wrong is what makes that honest: a rate of zero means "nobody filled in which document", not "your knowledge is fine". It counts REPORTS, not open flags: a document you fixed last week still shows the answers it broke, because it broke them. The definition field ships the exact arithmetic. The read is bounded; truncated: true means there were more confirmed-wrong reports than one request folds.

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/review-flags/analytics" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbReviewAnalyticsResponse.

Exemple de réponse

{
  "confirmed_wrong": 7,
  "documents": 2,
  "items": [
    {
      "agents": [
        "front-desk"
      ],
      "categories": [
        "outdated"
      ],
      "confirmed_wrong": 3,
      "document_id": "doc_4b1e77a0",
      "share": 0.6
    }
  ],
  "knowledge_fault_rate": 0.7142857142857143,
  "linked_to_document": 5,
  "scanned": 7,
  "truncated": false
}

POST /v1/kb/search#

Search your knowledge base. Searches your **published** documents and returns ranked passages with the matched words marked in a snippet. Natural-language questions work, the query is matched both by meaning and by word, so a Greek question finds a Greek document written with or without accents. filters.lang is the language of the person searching, not a hard filter: a document with no version in that language still answers, in the language it has. filters.collections, filters.tags, filters.categories, filters.audience and filters.source_authority narrow; nothing here can widen a search beyond your own published content, and a document outside its validity window never appears. filters.audience and filters.source_authority are the standardised taxonomy (see the create call). Both are lists and both match ANY of the values you send, so "source_authority": ["official", "verified"] is "trustworthy material only". **A document that declares neither is not returned by either filter**, absence is not a wildcard, and "nobody labelled this" must not be answerable as "this is official". An authority outside the four is a 400, not an empty result set. One thing to know about timing: the labels are stamped onto a document's passages when you PUBLISH it. Re-labelling an already-published document does not take it out of search, and it does not change what these filters match until you publish again. **This call is rate limited**, it embeds your query, which is a per-request cost, and the allowance comes from your plan. A refusal is 429 with retry-after; the current allowance is on every response as x-ratelimit-limit. A search that matches nothing is a normal 200 with outcome: "zero" and no results. Those queries are what the zero-result list in GET /me/search-stats is built from, they are the questions your customers asked that your content does not answer yet.

Corps de la requête

application/json Schéma : KbSearchRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/search" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "filters": {
    "audience": [
      "customers"
    ],
    "collections": [
      "col_9f2a1c7d"
    ],
    "lang": "el",
    "source_authority": [
      "official",
      "verified"
    ]
  },
  "limit": 5,
  "query": "πόσο κοστίζει το λεπτό ομιλίας;"
}'

Une réponse réussie renvoie un KbSearchResponse.

Exemple de réponse

{
  "count": 1,
  "outcome": "hit",
  "query": "πόσο κοστίζει το λεπτό ομιλίας;",
  "results": [
    {
      "chunk_id": "doc_4b1e77a0#0",
      "collection_id": "col_9f2a1c7d",
      "document_id": "doc_4b1e77a0",
      "highlights": [
        {
          "end": 51,
          "start": 41
        },
        {
          "end": 66,
          "start": 52
        }
      ],
      "lang": "el",
      "rank": 1,
      "score": 0.0328,
      "section": "Τιμολόγηση",
      "snippet": "…Η χρέωση γίνεται ανά λεπτό ομιλίας…",
      "title": "Τιμολόγηση 2026"
    }
  ],
  "took_ms": 41
}

POST /v1/kb/search/selection#

Record that a result was opened. Counts a search result the person actually opened. It is what turns the search report's selection rate into a real number, without it, "how often does a search lead anywhere" cannot be answered. Send the query exactly as it was searched; we key the count by the query, not by the document.

Corps de la requête

application/json Schéma : KbSelectionRequest

Réponses

Exemple de requête

curl -X POST "https://api.voiceland.ai/v1/kb/search/selection" \
  -H "Authorization: Bearer VL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "document_id": "doc_4b1e77a0",
  "query": "πόσο κοστίζει το λεπτό ομιλίας;"
}'

GET /v1/kb/suggest#

Autocomplete and did-you-mean. Completions for what someone is typing, drawn from your own documents' titles and headings, your tags and categories, and the searches that actually found something. When nothing completes the prefix, did_you_mean offers the nearest known word, accent-insensitive, in Greek and English. It answers from memory and never searches, so it is safe to call as the user types. The first call after a deployment can legitimately have nothing to say: ready: false means the word list for your account is still being built, and it is not an error.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/kb/suggest" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbSuggestResponse.

Exemple de réponse

{
  "ready": true,
  "suggestions": [
    {
      "kind": "title",
      "score": 3,
      "text": "τιμοκατάλογος 2026"
    }
  ]
}

GET /v1/me/search-stats#

Search report. What people searched for, whether they found anything, and whether they opened it, per day, plus the top queries and the **zero-result list**. The zero-result list is the content-gap report: every question your customers asked that your published content could not answer. It is the shortest route from "the agent said it doesn't know" to knowing what to write next. Defaults to the last 30 days. truncated is true when the window held more distinct queries than one request reads, narrow the window rather than raising the limit.

Paramètres

Réponses

Exemple de requête

curl "https://api.voiceland.ai/v1/me/search-stats" \
  -H "Authorization: Bearer VL_API_KEY"

Une réponse réussie renvoie un KbSearchStatsResponse.

Exemple de réponse

{
  "from": "2026-07-01",
  "to": "2026-07-31",
  "totals": {
    "hits": 1102,
    "queries": 1284,
    "selection_rate": 0.498,
    "selections": 640,
    "zero": 182,
    "zero_rate": 0.142
  },
  "zero_results": [
    {
      "count": 37,
      "outcome": "zero",
      "query": "επιστροφη χρηματων"
    }
  ]
}

Schémas#

Error#

Error envelope returned for non-2xx responses.

KbChunkPreview#

KbChunksResponse#

KbCollectionPatchRequest#

KbCollectionRequest#

KbCrawlScope#

KbDiscoveryResult#

KbDocumentPatchRequest#

KbDocumentRequest#

KbDocumentsResponse#

KbImportInlineRequest#

KbImportInlineResponse#

KbIngestResponse#

KbJobsResponse#

KbObjectImportRequest#

KbObjectSourcePatchRequest#

KbObjectSourceRequest#

KbObjectSourceView#

KbObjectSourcesResponse#

KbReviewAnalyticsResponse#

KbReviewDismissRequest#

KbReviewFlagsResponse#

KbSearchFilters#

KbSearchHighlight#

KbSearchRequest#

Search your published knowledge.

KbSearchResponse#

KbSearchResult#

KbSearchStatsDay#

KbSearchStatsQuery#

KbSearchStatsResponse#

KbSelectionRequest#

KbSuggestResponse#

KbVariantPatchRequest#

KbVariantRequest#

KbVariantResponse#

KbVariantsResponse#

KnowledgeCollection#

A shared container of documents. Agents attach collections by id; many agents may hold the same one, and attaching never writes to it.

KnowledgeDocument#

One piece of knowledge. Only a published document inside its validity window is searchable.

KnowledgeDocumentFaultRow#

KnowledgeDocumentVariant#

One language of a document: title and text only.

KnowledgeJob#

KnowledgeJobProgress#

KnowledgeRecrawlState#

KnowledgeReviewFlag#

KnowledgeSource#

Where the content came from: typed in (inline), uploaded (upload) or fetched (url). Kept as provenance for citations after a redaction, even though nothing re-fetches it then.

KnowledgeUsageStats#

Suggestion#