forge_qa
Q

The support desk
that thinks.
You confirm.

A platform for those who provide support on software products. The AI classifies, searches and answers from the manuals. The operator confirms with one click.

3
Levels
1
Click
N
Products
Corpora
01

From question to answer

The path of every request is explicit and traced: assigned category, confidence reached, sources cited.

A vendor’s knowledge lives in the manuals. The questions do not. Every product carries a corpus of procedures; those who support the customers repeat the same answers, searching by hand. forge_qa puts that corpus to work, whatever the customer base.

01

Frames

Classification with a lightweight model; category correctable by hand.

02

Searches

Knowledge base queried by semantic similarity; sources visible.

03

Decides

Confidence score: above threshold it answers, below it hands off.

04

Confirms

One click and the answer becomes official, with audit.

The final confidence is the product of two independent factors: the one declared by the model and a factor derived from the relevance of the retrieved passages. Below 0.45 human queue · from 0.45 automatic answer · 0.70 high confidence in the analytics.
02

Confirm AI answer

Previously, ratifying an automatic answer cost three steps. Now it is one click: the generated answer becomes the official answer instantly, with operator and closing timestamp recorded for audit.

Direct transition

From automatic answer to resolved, atomic.

Single source

The confirmed text is the answer sent to the client.

Traceability

Who closed it and when, always recorded.

request · detailClient · Delta S.r.l.
questionHow do you restore access for a user who has lost their credentials?
generated answer
From the administration panel, open the user’s record and start the reset: a secure link with an expiry is generated…
final confidence · 0.82 · sources · manual §3.2, §3.5
Confirm AI answer
Demonstration example. The button is available to those who can manage the request.
03

Knowledge enters like this: upload the manuals, index them

The knowledge base is not programmed, it is fed. From the console you register a product source, upload the documents and re-index: from that moment the product knows how to answer, citing what it has read.

01

Register the source

One document container per product: watertight compartments, no cross-contamination between scopes.

02

Upload the documents

Manuals, procedures and notes enter from the console. No code changes required to add knowledge.

03

Re-index

Indexing in delta — changed documents only — or full. One click and the corpus is up to date.

04

Verify

A test page shows the health of the base and lets you query it before putting it in front of clients.

When the base has no relevant context, the system says so and does not fabricate: the request passes to human handling instead of producing an arbitrary answer.
04

Every confirmed answer teaches the next one

The knowledge base does not live on manuals alone: every resolved request generates a candidate “question / answer” entry. Entries become knowledge only after human review.

01

Collection

The entry is born at the moment of resolution, one per request, without ever slowing the operation.

02

Review

Editable text, inclusion entry by entry. The preview is the document that will be published.

03

Publication

Approved entries enter the product source and it re-indexes. A new draft opens.

04

Clean corpus

Excluded entries stay in the history. The review keeps weak answers out.

A governed loop: what gets published becomes the source of future answers — the human step is what keeps the corpus standard high over time.
05

Everything a support department needs

triage

Classification

Every request is framed into categories with a lightweight model; the category stays correctable.

retrieval

Answers from manuals

The document base answers citing sources, with a confidence score on every answer.

gating

Confidence threshold

Above threshold the answer is automatic; below it passes to human handling.

1-click

AI confirmation

A good answer closes with a gesture: it becomes the official answer, with audit.

workflow

Request lifecycle

Ownership, resolution, cancellation and reopening: explicit states with history.

tenancy

Multi-tenant

Software houses, resellers and clients, with provisioning and per-tenant isolation.

catalog

Products and entitlements

Each product has its corpus. Reseller cap and whitelist decide who sees what.

finops

AI costs

Tokens and cost tracked per call and per model, with hierarchical spend quotas.

gdpr

Data privacy

Client master data is never sent to the AI; fonts and resources self-hosted.

security

Security

Anti-CSRF, TLS, Argon2id hashing, client data never sent to the AI.

06

One platform, many houses in white-label

forge_qa is multi-tenant by construction. The reseller resells to its clients what the software house grants; each level sees only its own.

L4

Software house

Governs the product catalogue, the corpora, the thresholds and sees the entire system. It is the “forge” brand.

L3

Reseller

Resells to clients within its own product cap; manages and confirms requests.

L2

Client

Submits requests choosing the product; inherits the reseller’s products, unless restricted.

07

What happens, measured as it happens

The dashboard reads the real state of requests: how many found an answer on their own, how many required a person, how confidence distributes.

Confidence distribution

High ≥0.70
Medium
Low
N/A

KB coverage per product

Prodotto A
Prodotto B
Prodotto C
Prodotto D
Automatic vs human, confidence distribution, base coverage per product: demonstration values, illustrating the kind of reading offered.
08

Every token has a price, and a cap

Artificial intelligence is a variable cost. forge_qa makes it visible and governable: measured per call, attributed per model, capped per tenant.

Tracked per call

Every call logs the tokens and the estimated cost, with the model that produced it.

Price list per model

Each model is assigned its price per million tokens: the cost of every request becomes a real number.

Hierarchical quotas

A monthly spend cap per user or tenant, cascading to descendants.

Zero cost when not needed

If the base has no relevant context, the generative model is not called: no expense.

Where the cost concentrates · per model

Q&A · Pro
Classification · Lite
Embedding

Generation with extended reasoning is the dominant line; classification and indexing weigh a fraction. That is where the optimisation lever applies.

09

One corpus per product

Products are in watertight compartments: every question is routed only to the corpus of the chosen product. The model works for any documented product, in any sector.

01 · corpus

Isolated base

Each product has its own document base: no cross-contamination between different scopes.

02 · tetto

Reseller cap

Defines which products a reseller can offer to its clients.

03 · whitelist

Client restriction

Further restricts what the individual client sees, within the granted cap.

04 · routing

Routing

Every question reaches only the corpus of the chosen product, with its sources.

10

Serious where it counts: data, access, continuity

forge_qa handles client data and their requests. The security posture is not an accessory: it is part of the product definition.

A1

Multi-tenant isolation

Every reseller and client sees only its own data; the hierarchy is enforced at model level.

A2

Robust credentials

Passwords with Argon2id hashing, forced change on first access, signed sessions.

A3

CSRF protection and TLS

Anti-CSRF middleware on every write, encryption in transit, secure cookies.

A4

Always-on service

The platform restarts on its own after any interruption; the database stays intact.

A5

Data outside the AI

Only the question text reaches the AI: no client master data.

A6

No CDN dependency

Fonts and resources self-hosted, in line with a data-conscious posture.

11

Support stops repeating itself. It starts answering.

forge_qa is an AYXZA system, in white-label, without giving up client data — whatever customer base you serve.

Resolved