Foreign workers employed through Israeli manpower corporations can now get answers in their own language. They follow guided scripts, ask free-text questions to a RAG assistant, or send a voice note. When the bot can't help, it opens a ticket that the Hebrew-speaking area manager reads in Hebrew, and the manager's reply reaches the worker translated back.
Manpower corporations employ foreign workers who speak Thai, Hindi, Sinhala, Russian or English. Their area managers work in Hebrew. Simple questions about a payslip, housing, insurance or a visa went by phone or messaging app. Nothing tracked them, and someone often had to translate by hand. The client wanted a self-service channel that answers the common questions in the worker's language and escalates the rest as a ticket that stays traceable in both languages.
The diagram is wider than the screen: swipe sideways to see all of it →
Every translated entity is stored once per language, together with a hash of its source text. The read path never calls the LLM, so a Thai worker opening a script gets stored Thai text with no added latency.
Step through one worker conversation. The content is illustrative and contains no client data. The pseudo-calls show what each stage does.
The model may answer only from the retrieved snippets, or reply with an exact __NOT_FOUND__ sentinel. Only then does a general-knowledge answer run, and it is stored with source=llm. Every answer logs its source, top score, retrievals, citations, tokens and latency. Trade-off: two LLM calls on a miss.
Scripts, nodes and FAQs are translated when an admin saves them and stored per language with a SHA-256 of the source. Items whose hash hasn't changed are skipped. Reading costs no LLM calls. Trade-off: more rows, and a few seconds before a new edit shows up in every language.
Saving FAQs returns as soon as the database commits. Indexing and translation run in their own DI scope in the background, with capped parallelism (10 embeddings, 5 translations). A failed job writes nothing, so the next save retries it naturally because the hash still differs.
A system-level master script and FAQ set is cloned into each new corporation, translations included, and tracked by an activation table. Corporations can customise their copy. Trade-off: a sync path is needed when the master changes.
Rich-text answers are translated in an HTML mode that keeps every tag. Plain titles go through a text mode that strips stray markup. If the content filter blocks a text, the original is kept and the save still succeeds.
The first retrieval design relied on semantic re-ranking, but the search tier in use didn't support it, so queries failed.
Switched to hybrid retrieval: a simple keyword query plus an HNSW vector query, top-8. The semantic configuration stays in the index for when the tier allows it.
Editing one FAQ in a list of 50 re-translated all 50 into every language: 400 LLM calls for a one-line change.
Added a SourceHash per entity and language. Unchanged items are skipped, so the same edit now costs 8 calls.
The save request waited for every translation and embedding to finish, and the admin UI hung while it did.
Batched the database upsert into one call, then moved indexing and translation into scoped background tasks with semaphores. The response returns in about a second.
After a deployment change, chat calls failed on an explicit temperature setting that the model no longer accepted.
Removed per-call sampling settings and made the deployment name the only model switch. Translation consistency now comes from strict prompts.
Chat and ticket times showed in the wrong time zone. EF Core returned DateTimeKind.Unspecified, and an earlier switch to local time made it worse.
The server stores UTC and marks it explicitly in DTOs. The client parser accepts strings without a Z and formats them per locale.
Onboarding a corporation copied the master script one node at a time, and parent links had to point at IDs that didn't exist yet.
Rewrote the copy in two passes: insert all nodes in one batch, build an old-to-new ID map, then fix parents in a second batch. That is 2 round trips per script, with translations copied too.
Some workers send voice notes instead of text. The manager saw an audio file with no context.
Added recording in the web app and Whisper transcription on the server. The ticket view interleaves each audio file with its transcript at the right point in the conversation.
Free scoping call · 30 minutes · No commitment
Write a few lines: the process or system, what hurts today (time, errors, cost or an AI pilot that stalled) and who uses it. After a short call you get a written plan and a fixed quote within ≈3 days.
Opens a pre-filled email · reply within one business day
The model reads. The code counts.
All projects