Cross-language knowledge-base search should help a visitor find approved material with the words they naturally use. It should not turn a single-language source into a promise that every translation, interpretation or local commercial detail is automatically correct.
Define success as retrieval, then test the answer
There are two related jobs. Retrieval finds the relevant approved material. Answer generation expresses that material in response to the visitor’s question. A successful retrieval does not prove that every sentence of the resulting answer is correct; an answer that sounds fluent does not prove that it came from the best source. Test both steps separately.
RobiFox’s public language description states that cross-language retrieval was tested across 20 languages and explicitly separates that claim from answer accuracy. Its published help benchmark documents a different test: 52 Hungarian questions, not a multilingual answer-quality test or the methodology for the 20-language retrieval claim. For your own deployment, evaluate retrieval and the resulting answers separately using the process below.
Build a multilingual retrieval test set
Choose a source page for each test before you run it. For every priority language, write a direct version, a natural visitor version and one version that uses local terminology. Include questions with an answer, questions that require a condition, and questions the knowledge base cannot answer. Do not use machine-translated copies only; ask local colleagues to add phrasing visitors actually use.
| Record | Why it matters |
|---|---|
| Visitor wording and language | Shows whether real phrasing, not only formal translation, finds the source. |
| Expected source | Lets reviewers distinguish a relevant match from an accidental one. |
| Retrieved candidates | Reveals whether the source was missing, ranked too low or simply ignored. |
| Answer and citation | Checks that the final response stays faithful to the retrieved material. |
| Verdict and repair | Prevents a retrieval defect from being “fixed” by adding duplicate content. |
Diagnose a failed search before adding content
When a visitor gets no answer, first inspect whether search found candidates. If it found none, the information may truly be missing, or the source may be too sparse to index usefully. If it found the right answer but still failed, the problem may be retrieval ranking or missing searchable wording. If it retrieved an outdated or conflicting source, fix governance before you add another answer.
RobiFox’s missing-knowledge workflow makes this distinction explicit. When candidates exist, an operator can link the visitor’s wording to the existing knowledge so it is searchable after re-encoding. When no candidates exist, the operator writes a new answer and teaches it to the assistant. The interface warns that adding a new fact where the existing one is already correct can create duplicates and conflicts.
Use wording variants as evidence
Maintain a small list of aliases and phrasing that produced a false negative. “Invoice,” “VAT receipt” and a locally used tax term may all mean the same visitor need, but they should only be attached to a source that truly covers that need. Document the source owner and change date. This keeps later editors from treating search vocabulary as a second, hidden policy.
Handle translation-sensitive content carefully
Some material can be safely paraphrased; some cannot. Prices, currencies, contractual clauses, compliance statements, technical limits and warranty exclusions deserve a stricter check. RobiFox says it does not convert prices or currencies in a foreign-language answer. Adopt the same discipline in your tests: check whether the original amount, unit, date and condition survive the answer.
For legal or regulated text, define the approved market version and use a human route for interpretation. Search can retrieve the policy, but it cannot decide whether a visitor’s individual situation qualifies for an exception.
Measure failures that change work
Report separate counts for genuine gaps, failed retrieval of existing knowledge, stale or conflicting sources, and questions correctly routed to a person. The first category calls for new content; the second needs better searchable wording or retrieval tuning; the third needs source ownership; the fourth may be the intended result.
RobiFox groups unanswered questions by frequency and retains the original wording, so teams can prioritise work by impact. A conversation log can also show which candidates and sources were involved. Use the feature set as a reference for the available controls, then set your own targets around visitor needs.
Cross-language search holds up when it can show why a result was found, why an answer was withheld, and what a team changed to improve the next search. For a complementary operating view, see multilingual support chatbot control.
Cover image: AI-generated editorial illustration, not a product screenshot.
