Knowledge holds the policies, explanations and business information Tephlo can retrieve for an answer. Keep this material current alongside your catalog, configured business facts and any enabled integrations.
Open Content → Knowledge in the console and upload a file, or paste text directly for something short like a returns policy or a price list caveat.
Upload: .pdf .docx .xlsx .csv .txt .md — choose a file or drag it onto the panel · up to 5 MB per file · or paste raw textEach source moves through PENDING and PROCESSING while it is parsed, split into passages and indexed, then becomes READY. Processing readiness is separate from approval and serving eligibility. The page polls while processing is in flight.
When a source fails, the reason is stated plainly rather than hidden:
PROCESSING indefinitely: once the retries are spent it is marked FAILED with this reason — at the latest within about an hour — and re-uploading starts a fresh revision.Re-uploading a source does not leave the old version half-live. A new revision is indexed alongside the current one and switched over in a single step, so a customer never gets an answer stitched from two versions of the same document. Until the switch, the current version keeps answering -- and if the new upload fails, it goes on answering while the failure is shown.
The filename decides whether an upload replaces anything. A document is identified by its name, so uploading a revised price list under the name it already has makes a new version of that document and the older prices stop being used. Giving it a new name instead -- prices-february beside prices-january -- adds a second document, and the first one is not retired: both stay live and both prices stay on file. Where two live documents describe the same article differently -- a price, a stock figure, whether it is available at all, a currency, a category -- the assistant stops stating the disputed detail and offers to confirm with your team, rather than choosing between two of your own documents. A price the two agree on is still given; it is the disagreement that is withheld, not the whole article. The upload report on the document says which articles are affected, which detail differs and what the other document says, so the fix is to re-upload under the existing name or delete whichever document is out of date. Better still, the upload check catches it first: when most of a file’s articles are already in one live document, it says so and offers to replace that document with this file, so the newer version simply takes over.
Select the source purpose offered by the uploader: business information, a product catalog or automatic classification. Inspect the detected purpose and product count before approval. Use Approve this content with a reason to bind your review to the exact source revision and content. Quarantine material that should stop being used.
The page distinguishes Not approved, Approved, Changed since approval and Quarantined. Re-uploading can invalidate an earlier approval, so review the new content. A verification stamp for freshness is a separate check from content approval.
Where the platform publication boundary is enabled, current approval and validity determine whether a processed source can serve. Check the current source state and workspace availability instead of assuming that upload or indexing publishes it. Content Home helps find outstanding review work across your content.
A customer’s question is not matched against your documents by keyword alone. Four things happen, and each exists to fix a specific way the naive version fails:
The bottom of Content → Knowledge has a question box that runs the same retrieval search the assistant runs. Type a question the way a customer would ask it and Search shows the five passages that came back, each with its document, the text that matched and a score. When the result says it was ranked by words only, the semantic half of the search was unavailable at that moment and the order is lexical.
A question you want to keep can be saved as a check. Expect this source on a result saves the question with that document as what the search must return; Expect a phrase saves it with words the returned passage must contain, matched without regard to case. Add up to five paraphrases — other ways a customer might word the same question — and each one is searched too. A check passes only when every wording finds what it expects.
Run all checks runs every active check and shows how many were found, the mode the search ran in and, for each check, the passages that came back with the rank at which the expected one appeared. Rank 1 means it was the first passage. A check found only at rank 4 or 5 is fragile: a longer question or a busier knowledge base can push it out of the top five. Pause a check to keep it without running it; delete it when the question no longer applies. The twenty most recent runs are kept.
A price list from 2023 is worse than no price list. Every source can carry a validity window and a human verification stamp, and the assistant treats them as a trust contract rather than as metadata.
Three rules follow from that, and they are the reason to bother setting dates at all:
The failure mode is not usually a missing document. It is a document that a human could interpret and a retrieval system cannot.
A catalog is a dedicated structured section of Knowledge, separate from articles and guidance. Prose retrieval cannot reliably answer “the cheapest one under 50,000 that is in stock” or supply an item to a cart. A validated product row can. Add products in Business → Catalog, or import a supported spreadsheet through Knowledge.
Header names vary between businesses, so common alternatives are recognised automatically: SKU, Product ID or Item ID; Product name, Name or Title; Final price, Sale price or Price; Stock quantity, Qty or Availability. The customer-facing price is preferred over the list price where a sheet carries both.
Aliases. An optional Aliases column (Also known as, Other names and Autres noms are recognised too, and any other heading can be mapped to it) holds the words your customers actually say for a product that its catalog name does not carry: a colour (“purple” for a phone the sheet calls Violet), a nickname, the word in another language. Separate several with commas. The assistant matches these words as strictly as it matches the product name — the column widens what it recognises without loosening exact matches, so a product named precisely is still answered, not offered alternatives. The words are for finding the product, not describing it: the product details the assistant presents to a customer never include the Aliases column. Aliases can also be edited product by product in Business → Catalog, without re-uploading a sheet.
Catalog answers are exact rather than paraphrased. The assistant filters and sorts by price, stock, category and brand against the real rows, so a price it quotes is a price you uploaded.
Only products belonging to the active, fully-indexed version of a catalog are visible to a customer. A half-imported spreadsheet cannot leak into an answer.
A row with the required SKU and commercial fields can also support a saved cart and a workspace-reviewed potential-order request. This still does not place an order, take payment, reserve stock or start delivery. See catalog and saved carts for the complete workflow and its boundaries.
With customer memory on, durable facts about a customer — their preferences, context they have already given you — are carried into future conversations, so they are not asked the same thing every time they write.
This is deliberately separate from the short-lived working memory that runs inside a single conversation. The quantity someone wants today is scratch for today; their preferred language or their usual delivery area is a fact about a person. Sensitive fields — card numbers, PINs, one-time codes — are refused outright and cannot be stored by either mechanism, whatever a configuration asks for. See working memory for the in-conversation half, and consent and retention for the controls.