Capability Details

Vertical LLM

Turn industry knowledge into a generative knowledge base so the model speaks the jargon and produces finished output.

01 / Technology decisions

Three trade-offs you cannot avoid

Each has its own cost; here is which side we chose and why.

How to build the knowledge base: structure first, or dump documents

Common practiceDumping PDF/Word files straight into the vector store

AlternativeExtract structure first, then load into the vector store

Our approachMix them: structured fields for lookup, body text for semantic recall

Dumping documents works well in a demo and collapses in production — it cannot answer structured questions such as “what is the scope of clause 4.2”. We settled on a hybrid: tables and clauses are first extracted into structured fields (exact matching), body paragraphs go through vector recall (for fuzzy questions), and the two result sets are merged and ranked. The cost is an extra extraction and cleaning pipeline; the payoff is accuracy that moves from guessing to lookup.

How to control generation: free-form writing or a rigid template

Common practiceLet the model generate freely

AlternativeFill in strictly to the template, no improvisation

Our approachTemplate skeleton + model fills the content, double validation on key fields

Free generation produces different output for the same input, and users will not trust it. A rigid template simply fails when it meets a case it does not cover. Our approach splits the document into skeleton and flesh: the skeleton (section order, mandatory clauses) is fixed by the template, the flesh (specific descriptions, examples) is generated by the model, and the result then passes a field validation that triggers regeneration if it fails.

Deployment choice: public API or local model

Common practiceCalling public LLM APIs: best results at the lowest cost

AlternativeOpen-source model deployed on-premise, data never leaves the network

Our approachDual track: local models on the intranet, API on the public internet, one interface

The first question from almost every client is whether the data can stay off the public internet and inside their intranet. On-premise deployment is therefore not an option but an entry requirement. The cost of a local model is that you handle quantization, inference acceleration and GPU memory planning yourself — a 7B model runs on a consumer GPU, but it chokes as soon as concurrency rises, so batching and queues are needed.

02 / Option comparison

One problem, three approaches

There is no optimal solution, only solutions that fit the scenario.

Solution	Accuracy	Implementation Timeline	Maintenance costs	Who it suits
Prompt engineering + public API	High (strong models)	1–2 weeks	Low (pay-as-you-go)	Internal trials, non-sensitive scenarios
Local model + simple RAG	Medium (limited by retrieval quality)	3–6 weeks	Medium (someone must watch VRAM and concurrency)	Sensitive data, high content volume
Local model + structured knowledge base + hybrid retrieval	High (traceable sources)	2–4 months	Medium-high (requires data cleaning and evaluation)	Production environment, where answers must be accountable

Accuracy figures are self-assessments, not a standard metric; a real comparison has to be run on the client's own question set.

What we do

Industry Knowledge Base Construction and Cleaning
Document template engine (policies / emergency plans / registers)
On-Premise Deployment and Data That Never Leaves the Domain
Scenario-based generation: fill in the form, get the finished document

Discuss requirements

Delivery format

Deployment modelOn-Premise / Private Cloud / Intranet

Scope of deliverySystem + Documentation + Training

Maintenance supportRemote + on-site

Source Code DeliveryNegotiable

Related Cases

Proven systems

Client names are anonymized per confidentiality requirements.

In-House Product / Vertical LLM

Industry Compliance Document Generation Engine (In-House)

Background: Enterprises need a complete set of policies, emergency plans and ledgers drafted to national standards. The traditional route relies on consultancies: long lead times and high cost.
Solution: Vertical LLM application: enter company information, get a complete set of compliance documents, editable online and exportable to Word.
Technology: Industry knowledge base (structured national standard clauses) · document template engine · LLM generation · form-based interaction
Outcomes: Covers two national standard systems and generates the full document set in one pass; the company revises it to its own situation and uses it.

View all case studies

03 / Implementation path

How many steps from zero to launch

The order cannot be reversed: starting with write operations or multi-step orchestration makes errors very hard to trace.

Extract 200 real questions

Draw them from the client's existing consultation records, acceptance Q&A and fault tickets — do not make them up. These 200 entries are the baseline for every later evaluation; without them you are tuning blind.

Build a minimal viable knowledge base

Start by turning the three to five most important documents (standards, procedures, templates) into a structured knowledge base, and get the full chain — ask, retrieve, generate, cite sources — working end to end.

Set an evaluation baseline

Run those 200 questions through it, manually label each as correct, wrong or incomplete, and calculate a baseline accuracy. Without a baseline, every later optimization is guesswork.

Add data, tune retrieval

Add documents, adjust chunk size and tune fusion weights according to the type of wrong answer. This round typically lifts accuracy by 15–30 percentage points.

Launch + event tracking

Plug it into the real business flow and log retrieval hits and user feedback for every question. Going live is not the finish line — it is where the evaluation set starts growing.

04 / Key metrics

Verifiable Numbers

No adjectives; write the specifications.

Documents Covered by the Knowledge Base50–400 documents per project　Four categories: standards / procedures / templates / registers

Time per generation8–25 seconds　On-premise, 7B-class model, with retrieval

Answer Traceability Rate100%　Every conclusion cites the document and clause number

Cold-start period3–6 weeks　From sampling questions to a working end-to-end chain

05 / FAQ

What technical leads ask most often

These are questions we have actually been asked, answered without marketing spin.

Which model do you use?

No lock-in. On the open-source side we use the Qwen / Llama family, chosen to match the client's compute and language needs; where stronger capability is needed and outbound access is allowed, commercial APIs are an option too. Models are replaceable — the real work is in the knowledge base and the evaluation.

What accuracy can you reach?

It depends on knowledge-base quality and question type. Factual questions (look up a clause, look up a parameter) typically reach 85%+; inferential questions (does this case apply?) are lower. We recommend measuring on the client's own question set rather than quoting someone else's numbers.

Can we just use an off-the-shelf knowledge base product?

Yes, if the requirement is general document Q&A. But if it has to speak the jargon — output structured by standard clauses, cite specific clause numbers, produce documents on an industry template — off-the-shelf products do not match that format, and customisation is still needed.

How is it guaranteed that data never leaves the intranet?

Model, vector store and application all run on the client's intranet with no access to any public address. Acceptance can include a disconnection test: cut the external network and everything still works.

Is updating the knowledge base a hassle?

Adding, editing and deleting documents happens in the admin console with incremental index rebuilding, no downtime. Changing structured fields requires re-running the extraction pipeline, which is best batched quarterly.