JNS-JIA-01
JIA
Connects documents, tasks, knowledge and client workflows.
1 Overview
JIA is a retrieval-augmented assistant that answers questions over a client's own documents and task history, then acts on them through tools. Every answer is grounded in retrieved sources and every action is logged.
Follow a recorded run of JIA →2 Block diagram
- 01IntentClassify: question, action or both
- 02RetrieveHybrid search: vectors + keywords
- 03RerankCross-encoder keeps the best 8 chunks
- 04ToolsTask lookup and updates when needed
- 05GenerateAnswer with inline citations
- 06VerifyCheck every claim maps to a source
3 Pin-out
- QUERYtextNatural-language question or instruction
- DOCSpdf · md · htmlClient documents, chunked and embedded on upload
- TASKSapiTask and project state from connected tools
- ANSWERtext + citationsGrounded response with source references
- ACTIONStool callsCreated or updated tasks, logged for audit
- TRACEjsonFull step-by-step record of the run
4 Specifications
| Parameter | Value | Notes |
|---|---|---|
| Latency p50 | 1.8 s | end-to-end, first token at 0.6 s |
| Latency p95 | 4.2 s | — |
| Cost per answer | €0.004 | median, incl. retrieval |
| Context window | 200k tokens | — |
| Retrieval depth | top-40 → rerank → top-8 | — |
| Grounded answers | 96 % | internal eval, n = 250 |
| Languages | EN · DE · ES · FR | — |
5 Stack
- Models
- Claude (generation)Embedding modelCross-encoder reranker
- Data
- Postgres + pgvectorObject storage for source files
- Runtime
- TypeScriptNext.js route handlersBackground job queue
- Observability
- Per-step tracesEval suite on every prompt change
6 Limits
- Answers only from connected sources; says so when nothing relevant is found.
- Tool actions above a risk threshold need human confirmation.
- Scanned PDFs without a text layer are skipped until OCR runs.