Solution
The governed deep-research engine for work that has to hold up.
The mid-2020s AI research market has two open wounds: citations that cannot be trusted, and runs whose cost nobody knows until after they run. DOCSI.AI answers both with architecture, not promises — a dollar figure approved before execution and mechanically enforced at every billable call, and a citation grammar that makes fabricated references unreachable rather than merely unlikely.
The problem the market has
Two failures every serious buyer runs into
Citations you cannot trust
Independent testing of leading AI search and research tools shows wrong-source attribution in a large share of results, and fabricated bibliographic references in generated text are a documented, rising problem — including in published academic work. Every incumbent claims to ground its answers. None enforces that a citation can only point at something actually fetched and kept.
Runs whose cost you cannot predict
No consumer research product shows a per-run cost. Frontier APIs can spend several dollars per run with no per-run cap parameter at all. Subscription quota cuts have produced public trust wounds. A whole ecosystem of third-party budget-enforcement middleware exists precisely because the primary tools do not do this.
What DOCSI.AI does differently
Eight strengths, each verified in code
Approve a price, and the software keeps it
Before anything runs, you see a low/expected/high cost estimate and a per-cap budget-fit view. During the run, every billable call passes through a single reserve-before-spend choke point enforcing seven independent caps. No hidden pre-run spend, no runaway concurrent calls.
Closed-world citations — fabrication is structurally impossible
Reports may only cite source identifiers that exist in the persisted, curated source index for the run. Unknown IDs, bare URLs, and model-written References sections are stripped by deterministic code. The References section is regenerated exclusively from what was actually fetched and kept.
A guaranteed deliverable, even under exhaustion
A synthesis reserve is carved out at plan approval and execution can never spend it. When budget or wall clock runs low, the run flips to wrap-up-and-synthesize and ships the best cited report the evidence supports, honestly banneredĀ — not a silent model downgrade.
The audit trail is the storage format
Every run is a directory of append-only, per-line-fsynced JSONL — approved plan, spend ledger, events, sources, learnings, claim audit. Inspectable with grep and jq, resumable without re-paying, crash-safe by construction. Compliance and e-discovery fit out of the box.
A shipped QA audit + bounded repair
A free, read-only defect detector runs over persisted artifacts and surfaces citation-closure violations, broken charts, and structural defects. Optional repair is opt-in, budget-capped, and forbidden from rewriting the evidence trail.
Typed search verticals in one pipeline
Web, news, scientific papers, and patents as first-class routes with structured filters — year ranges, assignees, family deduplication — alongside pinned canonical documents and explicitly labelled model knowledge stored separately from fetched evidence.
Concrete, predictable economics
Effort presets from a $1.50 quick run to $45+ max-depth runs, with per-call spend logged to the cent. Resumable runs, source dedup, and automatic saturation stop mean you rarely pay for diminishing returns — and never pay twice for the same fetch.
Engineering maturity beyond a prototype
A high test-to-source ratio, executable architecture rules, a ratcheted lint-debt ceiling, three-provider LLM abstraction, and continuous paid end-to-end validation. A low-delivery-risk asset with a contract-grade service blueprint already written.
How the budget kernel works
Governance wired into every billable call
Budget governance is not a dashboard bolted on after the fact — it is a reservation-before-spend kernel every LLM, search, scrape, and OCR call passes through.
Estimate & approve
The plan is presented with a low/expected/high USD estimate and a per-dimension fit table showing which cap will bind first. Headless runs fail closed if the high estimate exceeds headroom.
Reserve-before-spend
Every external call — LLM, search, scrape, OCR — passes through one choke point. In-flight reservations count against caps so concurrent calls can never overspend.
Ledger & stop
An append-only, per-line-fsynced ledger records every committed operation. A deterministic escalation ladder warns, protects the synthesis reserve, and hard-stops within a small grace band.
Seven independent caps apply — total USD, per-provider native units, LLM USD, source count, and wall clock among them. A protected synthesis reserve guarantees the report is always affordable. Fully autonomous spending is an explicit non-goal — approval always sits with a human.
Closed-world citations
Fabricated references, structurally out of reach
A citation in a DOCSI.AI report can only point at a source identifier that exists in the persisted, curated index for that run. Unknown IDs, bare URLs smuggled into prose, and model-written References sections are stripped by deterministic code — and the References section is regenerated exclusively from what was actually fetched and kept. Model knowledge, when used, is quarantined into a separate "Background estimates" section and labelled as context, not evidence.
What this prevents
- Reports citing sources that were never fetched.
- Bare URLs and model-written References sneaking into the deliverable.
- Silent double-billing across resumed or retried runs.
- A polished PDF that hides which claims are single-source.
What it does not replace
- Misreading of a source's actual meaning by an LLM.
- Human judgement on whether a source is authoritative for your context.
- Regulated professional review — legal, scientific, medical, financial.
- The natural quality ceiling of the underlying rented models.
How it compares
Where DOCSI.AI stands alone
Compared against frontier hosted deep-research tools and open-source recursive researchers, the governance and evidence cluster is what no rival combines.
Simplified for readability. "Frontier hosted" covers the leading commercial deep-research products; "OSS recursive" covers open-source recursive research agents.
The honest counterweight
Where the incumbents genuinely lead
Speed
Frontier hosted tools finish in 5–10 minutes. DOCSI.AI's flagship runs take 15–150+ minutes because governance and evidence traceability add real work.
Output breadth
Hosted incumbents ship slides, dashboards, and 25+ language exports. DOCSI.AI focuses on Markdown, HTML, and a designed, branded PDF.
Private-data connectors
Frontier vendors offer MCP and workspace connectors into private data stores. DOCSI.AI does not — today.
Casual-user economics
A $20/month unlimited subscription undercuts any per-run story for casual research. DOCSI.AI's case is governance, not cheapest.
DOCSI.AI is not aimed at casual research. Its market is governed, scrutiny-facing research — precisely where citation reliability and cost control move from annoyance to disqualifier.
Who it's for
Built for teams that must defend their answers
Strategy consultancies
Compress the first one to two weeks of an engagement into an approvable, budgeted overnight run — with an evidence trail the partner can stand behind.
VC / PE deal teams & diligence
Claim-level trust labels, corroboration passes, and page-locator PDF extraction for data-room documents. Per-run cost is noise against deal economics; auditability is not.
R&D & IP scouting
Typed Patents + Scientific Papers + news + web in one governed pipeline, with family deduplication and disciplined scope language for prior-art sweeps and FTO pre-screens.
In-house market intelligence
A resumable workspace per topic. Monthly follow-up waves extend a landscape for the marginal cost of the delta; the ledger supports internal chargeback.
Research agencies & shops
Effort presets function as a price list mapping to tiered offerings. MIT-style local execution means client evidence never transits a third-party SaaS.
Economics
Predictable, not just low
Effort presets span from around $1.50 for a quick 15-minute run up to $45+ for a max-depth report spanning hundreds of sources. Per-call spend lands in a per-run ledger, sliced by API, node, model, or session. On BYO keys, that is fully attributable — no subscription floor, no quota cliff, no vendor margin on inference.
- Never pay twice — resume, cache, dedup, and a visited-URL registry.
- Automatic stop-loss when new-URL yield and per-node learning drop below thresholds.
- Per-role model routing lets you spend frontier only where it matters.
The right frame
Not "the cheapest research per token" — but the only pipeline where the price is agreed before the work, and the deliverable is a cited, inspectable artifact whose remaining human review is exactly what the audit trail is designed to make fast.
Honest limitations
Where the story ends, on purpose
Beta status, local-only today
No hosted service or HTTP control plane yet. A specified jobs-API blueprint exists, but is not shipped.
Command-line form factor
Adoption today assumes a technical operator. It suits consultancies, R&D groups, and agencies with one — less so terminal-averse analysts.
External API dependency
The pipeline rides rented LLMs, a search backend, and scrape providers. BYO keys mitigate lock-in but you inherit provider pricing, quota, and outage risk.
Bounded quality guarantee
Closed-world citations prevent fabrication, not misreading. Quote verification is best-effort. Expert review is still required for regulated decisions.
These bound the claim; they do not undermine it. The core differentiators are architecture, exercised on real paid runs. The honest sales motion: governed research for technically capable, scrutiny-facing teams — with a specified path to broader distribution as the product matures.
Put governed deep research to work.
Plan the research. Control the spend. Trace every citation.