Ask in plain language. Phronexa reasons over your trusted sources, answers with exact citations, and abstains rather than guess. Arabic-first, entirely on your ground.
- Plans & self-corrects
- Exact per-claim citations
- Abstains rather than guess
- Arabic-first
A reasoning loop that checks its own work.
Classic search fetches a few passages and hopes. Phronexa runs a governed loop: it plans, gathers, grades its own evidence, and only answers when the evidence holds. Otherwise it searches again, or abstains.
Understand
Reads the question in context, resolves what it’s really asking, and scopes the search to the right sources.
Plan
Breaks a complex or multi-part question into steps, instead of a single blind lookup.
Gather
Searches iteratively across every connected source with hybrid retrieval and reranking.
Grade & correct
Judges whether the evidence actually answers the question, and re-searches or broadens when it falls short.
Synthesize
Composes an answer strictly from the gathered evidence, in the language you asked.
Verify
Checks every claim against its source before a word is revealed, then regenerates or abstains if it doesn’t hold.
Cite
Ties each claim to the exact passage, down to the article and paragraph, highlighted in the original.
Abstain
Says “I don’t know” rather than guess when the evidence isn’t there, with a measurable bound on error.
Watch it plan, search, and cite in real time.
The answer streams only after it clears the grounding checks, so you never see an unverified claim take shape. Sources appear inline, tap to open the exact passage, and speak the answer aloud. The same experience on the web, in Teams, and on WhatsApp.
- Step by step: it plans, searches, and answers in the open
- Correct-before-reveal streaming, with no flicker of a wrong claim
- Inline sources you can open to the exact passage
- Voice questions in, spoken answers out
Speak to it. Hear it answer.
Ask out loud and Phronexa transcribes it. Get the answer read back in a natural voice. Both directions work in the chat and across channels, in Arabic and English, on your own speech models.
Speech to text
Ask by voice and it transcribes accurately, tuned for Arabic and English.
Text to speech
Answers read back in a natural voice, on the web, in Teams, and on WhatsApp.
Every claim opens to its exact source passage.
Answers aren’t vaguely “based on a document.” Each claim links to the precise article or paragraph it came from, highlighted in the original source, so a reviewer, auditor, or regulator can verify it in a click.
- Citations resolve to the exact passage, not just the file
- The cited article highlighted in its source PDF
- Logical locators: “Article 5”, page, and paragraph
- Provenance carried from ingestion through to the answer
Structured, reviewed knowledge behind every answer.
Documents don’t stay as raw text. They’re modeled into a reviewed knowledge base of entities, relationships, and typed chunks, so retrieval finds the right passage, every claim ties to an exact locator, and only approved content can ever answer.
- Documents modeled into entities, relationships, and typed chunks
- Human-reviewed and approved before it can answer
- Precise retrieval of the exact article, not a fuzzy match
- OCR, tables, and Arabic-first tuning built in
A measurable bound on being confidently wrong.
When the evidence is thin, Phronexa abstains rather than guess. The restraint is calibrated, so the chance a revealed answer is wrong stays under a limit you set. Restraint you can put in a contract.
- Abstains rather than fabricate when evidence is insufficient
- A provable statistical bound on revealed but wrong answers
- Every answer graded and tracked by an evaluation harness
- Declines out-of-scope questions instead of improvising
One search across every trusted source.
A source is anything the platform can read as evidence. Connect one and agentic search reasons over it, always scoped to who’s asking.
- Document knowledge bases
- Structured databases (NL→SQL)
- Knowledge graphs
- Spreadsheets & tables
- Web sources
- SharePoint / Microsoft 365
- Uploaded files & images
- Your REST APIs
Trust built into every answer.
Every answer inherits the same guarantees, so grounded knowledge work stays safe and accountable.
Grounded & cited
- Answers only from approved, reviewed content
- Exact citations, traceable to the passage
- The cited article highlighted in its source
- No answers from superseded or expired rules
Provable restraint
- Abstains rather than guess
- A measurable bound on revealed but wrong
- Verified claim by claim before reveal
- Graded and eval-tracked over time
Scoped to who’s asking
- Retrieval respects each person’s access
- A user only sees what they’re cleared for
- Real identity, never a shared account
- Federated across many knowledge bases
Traceable & audited
- Every question and answer recorded
- Attributable to a verified identity
- A tamper-evident, hash-chained trail
- The whole reasoning loop is traceable
Real answers. Real citations.
Sovereign, Arabic-first AI, ready to prove its value on-premises, in weeks.
Start with one high-value use case and measure it on real data.