Source · govern documents
NexSpec Hub
Specification and document master data hub
A single source of truth, a version status flow, source files that can't be overwritten, and SHA-256 chained audit.
Home Products Nex Knowledge Base
Product 02 · Know
Nex Knowledge Base
On-premises Knowledge search and Q&A
So that every file on your company's intranet can be found, cited, and trusted.
3.54s
End-to-end Q&A response
0.75s
Fuzzy search time
0
External RAG dependencies
Problem → Approach
Time spent finding a document should be spent on the work
The files aren't missing — the search is wrong. And "send the data off-network to search it" is a path R&D organizations can't take.
| The old way | The Nex Knowledge Base way |
|---|---|
| Files scattered across drives and shared folders | Unified ingestion, chunking, and indexing make the whole directory searchable |
| Finding one file means opening five or six before you hit it | Fuzzy search surfaces the matching passage so you can judge without opening |
| File names are searchable, but contents aren't | Keyword inverted index and three-path vector recall run in parallel, so paraphrases still hit |
| Body text in scans and PDFs can't be read | Structure-aware chunking covers paragraphs, clauses, and table rows; scans go through visual OCR |
| Can't risk feeding company documents to an off-network LLM | Retrieval and generation run entirely on the intranet, with no external RAG platform |
| No way to verify whether an answer is right | Every answer carries sources and relevance scores and links back to the source for checking |
Capabilities
Six modules covering view, find, ask, and govern
Organize workbench
Twelve-plus rows on one screen, with select-all-by-type and batch confirmation. It compresses one-by-one decisions into a few operations.
Fuzzy search
With highlighted matching passages and source categories. Judge whether it's the one you want before opening the file.
Knowledge Q&A
Answers from documents in the library, with numbered citations and reranker scores, letting users judge credibility themselves.
First-scan wizard
A four-step wizard carries you through the first round of ingestion — no need to grasp chunking and indexing first.
Snapshots & rollback
Takes a snapshot before organizing; changes roll back, with a diff-comparison view.
Adjustable retrieval parameters
Endpoint configuration, self-checks, and retrieval parameters are visible and adjustable in settings; reranking and embedding failures degrade automatically without blocking Q&A.
Workflow
The five steps behind a single question
STEP 01
Structure-aware chunking
No cutting by character count. Clause numbers, section headings, and table rows all inform boundary decisions.
STEP 02
Dual-path retrieval
Dense vectors (Qwen3-Embedding-0.6B, 1024-dim) and a full-text inverted index (Chinese bigram + English words, a BM25 variant) are recalled at the same time.
STEP 03
RRF fusion
k=60. It looks only at rank, not score scale, avoiding the instability that arises when cosine similarity and BM25 have different dimensions.
STEP 04
Reranker refinement
Qwen3-Reranker reranks candidate chunks. On failure it falls back automatically to pure hybrid retrieval, marking the degradation without blocking search.
STEP 05
Generation and source assembly
An on-premises generation model answers and stitches the cited sources back in, labeling numbers and relevance.
ALWAYS
Never leaves the intranet
Chunking, embedding, retrieval, reranking, and generation all happen on your intranet, with not a single external request in between.
Interface
Organize workbench: a dozen-plus rows in one screen
Organizing thousands of files can't mean reading page by page. We compress row height to 46px, move the rationale to a secondary line, and enable select-all-by-type batch confirmation — turning "N individual decisions" into "a few".
N
Internal Knowledge BaseFile tidying & search
Search files or ask a question…Ctrl K
Personal knowledge base
Organize workbench1,284 Fuzzy search Knowledge Q&A Snapshots & rollback3Shared knowledge base
Fuzzy search Knowledge Q&A SettingsQuick tips
Click the search box at the top, or press Ctrl K to jump to search anytime.
User guide
Everything is processed on this machine; file contents never leave it and never pass through any external service.
Organize workbenchAI suggests, you batch-confirm
Choose folder
yu-peng
All1,284
Rename26
Archive74
Duplicates39
Reclaimable3.3 MB
1,284 total · click "Details" to ask about this file
File / rationaleTypeActions
DOCXBody control module_function specification_v3.2.docxContent fingerprint match · 21.4 KB reclaimable
Rename
Rename
DOCXWindow anti-pinch function_detailed specification_v2.4.docxSimilar filename 96% · suggest completing the document number
Rename
Rename
XLSXForward collision warning_interface specification_v1.8.xlsxContent fingerprint match · 40.4 KB reclaimable
Archive
Archive
XLSXGateway diagnostics_communication matrix_v5.1.xlsxContent fingerprint match · 30.7 KB reclaimable
Archive
Archive
PPTXCenter display_interaction flow specification_v0.9.pptx92% body overlap with v0.8 · suggest keeping the newer version
Duplicates
Merge
PDFSensor calibration report_2026Q2.pdfContent fingerprint match · 257.0 KB reclaimable
Archive
Archive
Scanned document_wiring harness drawing_A3_007.pdfVisual OCR result · low confidence, needs manual review
Manual review
Open
XLSXDiagnostic trouble code_mapping table_v2.0.xlsxContent fingerprint match · 30.7 KB reclaimable
Archive
Archive
Body control module_function specification_copy.docx100% match with the primary file · suggest merging
Duplicates
Merge
Vehicle network topology description.mdContent fingerprint match · 4.8 KB reclaimable
Archive
Archive
6 selected
Clear selection
Confirm 6 items
Auto snapshot before execution
The value of finding a file isn't in "getting a hit"
It's in "knowing whether it's the one you want without opening it". So fuzzy search surfaces the matching passage directly, along with source categories — saving you those five or six rounds of "open, glance, close".
- Matching passages shown outright. The matching sentence shows right in the results list.
- Scores aren't hidden. Sources and reranker scores are shown outright, so you judge the confidence yourself.
- Mistakes can be undone. Both organizing and indexing keep snapshots, so you can roll back and diff.
Measured
Measured retrieval quality
Below is measured data from an on-premises environment. We'd rather give verifiable mechanism names and numbers than "significantly effective".
| Test item | Measured | Notes |
|---|---|---|
| Model endpoint connectivity | 3 / 3 online | All three endpoints — generation, Embedding, Reranker — are available |
| Index scale (sample) | 7 documents → 49 chunks · 1053 terms | 49-dimensional vectors, 1053 inverted terms |
| Fuzzy search "confidentiality clause" | 0.75s | Hit ranking 0.9949 / 0.9946 / 0.4300 / 0.1608 |
| Semantic search (synonym paraphrase) | 0.69s | Related document score 0.9937 |
| Knowledge Q&A | 3.54s | 2.86s including generation, answer with [1][2] citations, source scores 1.0000 / 0.9998 |
| Reranker self-check | Separability passed | Relevant min 0.0594 > irrelevant max 0.000048 |
Why RRF instead of weighted sum? Cosine similarity and BM25 have completely different dimensions, so weighting them directly destabilizes the result. RRF looks only at rank, which adapts naturally to heterogeneous retrieval sources.
Design Decisions
Built to be trusted, not to look impressive
Scores aren't hidden
Sources and reranker scores are shown outright. Letting users judge credibility themselves is more useful than a bare answer.
Degrade, don't block
On reranking failure it returns to pure hybrid retrieval; embedding failure degrades the same way, marking only the state — retrieval availability comes first.
No external platforms
No external RAG platform; the entire retrieval pipeline is self-built. Company data never leaving the intranet is the one thing we won't trade away.
Data Boundary
Data stays on-premises — this line doesn't move
Files, vectors, indexes, and logs all stay in your data center. Decryption keys, access tokens, and conversation history each stay where they belong.
- Files never leave the domain. Original files and extracted body text are both stored on intranet storage.
- Vectors never leave the domain. Embeddings are computed by an on-premises endpoint, and vectors are stored in the intranet index.
- Model endpoints stay local. The generation, embedding, and reranker endpoints are all deployed on-premises and run even with the network cut.
- Answers are verifiable. Every answer carries sources and relevance scores and links back to the original text.
- Permissions isolated by project. Retrieval scope is bound by project permissions; it never fetches data beyond your authority.
Who It's For
For teams that have already accumulated large volumes of internal material
Document-heavy R&D teams
Specifications, design standards, interface documents, and test reports accumulate over time, and search needs are frequent.
Organizations with hard data-egress constraints
Security policy doesn't allow handing R&D data to external SaaS or cloud LLMs.
Cross-border collaboration teams
Chinese, Japanese, and English documents coexist, and the same retrieval capability must cover all three languages.
Continue
The other two links in the chain
Test · turn it into an asset
NexTest Studio
AI test closed-loop platform for automotive software
Atomic requirement extraction, test specification generation, end-to-end traceability, and change impact analysis.
All · the full picture
Back to home
How the product line connects into one chain
Document master data → internal knowledge → test closed loop, with one single fact carried all the way through.
Request a Demo
Bring a real question and ask it
We'll demonstrate semantic query rewriting, answer citations, and source-score verification live.