How Deep Synthesis Works

What Deep Synthesis is

Deep Synthesis takes a legal question or a fact pattern, written the way you would actually phrase it, and returns a structured, cited legal-research report. The product describes itself plainly as "Deep Synthesis for Philippine Law," and that scope is deliberate: its databases, statute index, case-law corpus, and prioritized web sources are Philippine.

What sets it apart from a general chatbot is not the writing at the end — it is everything that happens in the middle. A chatbot answers from memory. Deep Synthesis instead retrieves, reasons, and verifies: it searches real sources for each query, reads them, builds the analysis from what it found, and then checks its own citations against those sources. The result is a Markdown report organized in IRAC structure (Issue → Rule → Application → Conclusion), with case-law depth and a References section whose links point back to the underlying documents.

One honest framing up front: Deep Synthesis is a research assistant, not legal advice. The report carries that disclaimer, and so does this article.

What you do — the six steps you see

From your side, the experience is six steps:

1. Describe the legal issue. A single input box. Write a question or a fact pattern in plain English. The product hint is candid: "cite specific statutes, doctrines, or fact patterns for best results."
2. Review and choose sub-questions. Deep Synthesis breaks your input into discrete legal sub-questions and shows them to you. You select which to research, add context (case facts, jurisdiction, time period), or refine the query if the breakdown is off. (If there is only one sub-question, this step is skipped.)
3. Watch the research run. A live timeline shows each stage with plain labels and running counts — cases found, sources gathered. You can step away; you are asked not to close or refresh the tab.
4. Handle the occasional checkpoint. If the legal database is briefly unreachable, the tool pauses and asks whether to continue with web sources only or cancel — your call. A partial error produces a clear "results may be incomplete" warning rather than a silent failure.
5. Read the report as it is written. The report streams in live. When it finishes, the header reads "Research complete" alongside a faithfulness badge ("{N}% verified"); hovering tells you how many citations, if any, could not be verified.
6. Export and reuse. One click to Export PDF, Export HTML, or Copy text. Exports carry a footer noting the report was generated by the tool. If a run fails partway, you can resume from where it stopped without losing the work already done.

Under the hood: the eight stages

Behind those six steps is an eight-stage pipeline. Each stage is real and observable on the timeline — which is the point: the path from query to report can be inspected rather than taken on faith.

1. Query understanding & planning. Before searching, an AI model understands the request. It normalizes the query into clean legal English (handling natural-language input and fixing typos); expands Philippine legal shorthand (RA → Republic Act, BP → Batas Pambansa, PD → Presidential Decree, RPC → Revised Penal Code, FC → Family Code, VAWC → RA 9262, plus agencies like NLRC, DOLE, SEC, BIR); runs a content-safety intent check; decomposes the question into up to four focused sub-questions; generates targeted search queries; and extracts entities (G.R. numbers, party names, dates, statute references). The output is the research plan you review.
2. Authoritative legal-database retrieval. For each sub-question, in parallel, the tool searches the Intellegal legal database across both case law and statutes. It de-duplicates and score-ranks the hits; an AI model then reads each candidate and judges relevance — HIGH (directly on point), MEDIUM (related), or SKIP. A safety net force-includes heavily-cited landmark cases so seminal authority is not lost. It fetches the full text of the strongest results, and performs cited-law discovery: it looks at which statutes the retrieved cases themselves cite and pulls those in — surfacing governing statutes anchored in how courts actually applied them.
3. Query refinement. Before touching the open web, the tool looks at what the database already covered, identifies gaps, and rewrites its web-search queries to avoid duplicating what it already has — so the web step adds genuinely new material.
4. Live web research. A research agent works the open web in multiple rounds, deliberately alternating between searching and reading — it opens and reads pages rather than collecting snippets. It prioritizes authoritative Philippine sources (the official laws library, the Supreme Court site, the official gazette) and blocks low-value ones (social media, forums); a model selects which results are worth opening, a page-reading service fetches the full text, and the legally relevant content is extracted and de-duplicated across rounds. A sensible cap keeps it converging instead of looping.
5. Context distillation. Legal documents are long, so each source is compressed to its legally relevant essence — short documents pass through in full; long ones are split and the most relevant passages kept (original wording preserved). For known statutes, the tool uses article-level retrieval across an index of 80+ Philippine laws, surfacing the specific provisions that matter instead of dumping an entire code.
6. Evidence-coverage check. Before writing, the tool checks that every sub-question is actually supported by evidence. If it finds a gap, it launches a small, bounded number of targeted supplemental searches to fill it — so the report is not confidently written on thin evidence.
7. Report synthesis. The tool writes the report, streaming it to the screen. The analysis is structured as IRAC with real case-law depth and a References section, and is guided by an explicit evidence-weighting framework: how directly a source matches the issue, the authority and citation weight of a case, recency, consistency across sources, and cross-validation between database and web findings. Citation form is handled carefully — full case citation with G.R. number and date on first mention, statutes by Republic Act and section.
8. Faithfulness & citation verification. Finally, the tool audits its own report. It extracts every citation and matches each against the sources it actually retrieved; verified citations get clickable links, and citations that cannot be matched are stripped or flagged as unverified rather than left in silently. It reconciles the in-text citations with the References list, computes a faithfulness score, and shows it as a badge.

Seen in a live run

The same stages, captured as they happened on the query above:

At a glance

For a typical query the whole run finishes in roughly 5–20 minutes (depending on how complex the question is) across about seven reasoning steps — where the same first-pass research, done by hand, would typically take a lawyer at least a day. Treat the timing as typical, not a promise; a complex, multi-issue question can take longer.

Being honest about the limits

Two points matter for anyone relying on legal research.

First, the faithfulness check reduces and surfaces unsupported citations and grounds the rest in retrieved sources. It does not make the report perfect. The score can be below 100%, and when it is, that is shown to you plainly rather than hidden. Read it as transparency and self-checking — not as a claim of zero error.

Second, the scope is Philippine law, and the output is a research assistant's work product, not legal advice. The value is that the draft is checkable: every verified citation links to its source, so you can open the authority and confirm it before you rely on it.

Why this matters, by reader

Practicing lawyers and in-house counsel. Dual-source grounding plus cited-law discovery means first-pass research starts from actual statutes and how courts applied them, not unsourced prose — and the faithfulness score with clickable sources makes the draft checkable. The IRAC report and one-click export fit how memos and briefs are already structured.

Law students and faculty. The sub-question decomposition and the stage-by-stage timeline are instructive in themselves — they model how a research problem is broken down and pursued, and the References section reinforces the habit of citing real, checked authority. The transparent pipeline is a useful basis for discussing how AI legal research should be evaluated rather than treated as a black box.

Bar-exam candidates. IRAC-structured output anchored to Philippine statutes and case law matches the reasoning the exam rewards, and article-level statute retrieval helps pinpoint the exact provisions behind a doctrine. Plain-language input lowers the friction of phrasing an issue naturally.

Related reading

• Case Contrast — when the question narrows to how two specific cases differ.
• Contract Review — the same verify-before-you-rely discipline, applied to a document you upload.
• Obligations & Contracts — a worked overview of one doctrine Deep Synthesis is often asked about.

FAQ