The platform
Read once. Every question after is a lookup.
Your documents already contain a database. Talonic finds it: every field, organised once, queryable by every team, system and agent, with every value traced to its line.
Read for everything, not one schema. One client needed 75 fields. One read of 5,000 documents holds 52,937 reusable concepts. The surplus is what the next team queries.
Four phases, one confidence gate. Phase 1 fills cells from the database with zero AI calls. Once a cell reaches 0.7 confidence, nothing later overwrites it.
Every value points back to its line. Value, source region, confidence, phase and reasoning on every cell. Click it and see how it got there.
01 · Sources
Point it at anything.
No templates, no training data, no field list up front. Upload or sync, and Talonic classifies each document and captures what it finds.
- 25+ formats. PDF, Office, scans, images, email, CSV, XML, JSON and ZIP archives, unpacked with their folder paths.
- 529 document types in German, English, French and Spanish. Your own types sit on top; nothing fails, it lands in Unclassified.
- Connectors. Upload, Amazon S3, Google Drive, Notion and Microsoft Teams. Duplicate files are dropped at upload.
- System data too. Snapshots of your TMS or ERP become a source, with every change between two snapshots captured.
DOCUMENT ONTOLOGY
529 document types. Zero templates.
From Schedule K-1 to Bill of Lading (Ocean), from Notarial Deeds to QC Inspection Forms. Talonic understands the structure of every enterprise document type out of the box.
Schedule K-1, Form 1099-MISC, Form W-8BEN, Form 1040, Form 990, and 48 more
Purchase Order, Commercial Invoice, Pro Forma Invoice, Goods Receipt Note, Three-Way Match Report, and 48 more
Bill of Lading (Ocean), Bill of Lading (Inland), Air Waybill (AWB), House Air Waybill (HAWB), Sea Waybill, and 48 more
Notarial Deed, Master Service Agreement, Statement of Work, Non-Disclosure Agreement, License Agreement, and 48 more
Commercial Register Extract, Certificate of Good Standing, Certificate of Incorporation, Annual Return, Director's Report, and 48 more
Patient Intake Form, Medical History Questionnaire, Informed Consent Form, Discharge Summary, Operative Report, and 48 more
QC Inspection Form, Certificate of Analysis (CoA), Certificate of Conformance, First Article Inspection Report, Material Test Report, and 48 more
Insurance Application, Policy Declaration Page, Insurance Policy Wording, Endorsement / Rider, Certificate of Insurance, and 48 more
Property Deed, Title Report, Land Registry Extract, Survey Plan, Zoning Certificate, and 48 more
Employment Application, Offer Letter, Employment Contract, Background Check Report, I-9 Employment Eligibility, and 47 more
02 · The database
A database that compounds.
Every field from every document resolves into one registry. It is discovered from the documents first; output schemas are mapped onto it later.
- Same meaning, one concept. Vendor Name, Supplier Name and Company Name resolve to one field, the variants kept as aliases.
- Three tiers. Core, established and emerging fields, promoted as more documents confirm them.
- Master instructions. The best way to read each field is learned across documents and reused on every run.
- Nothing discarded. 1,042,665 source observations from one client’s first 5,000 documents, organised into 52,937 concepts.
03 · Specs
Output you control.
A spec is the shape a downstream system needs. Generated per document type, or written for Dynamics, Ivalua, TMW or any target.
- Versioned like code. Live, workshop draft and full history. Breaking changes are flagged before publish.
- Test before you ship. Run a draft against samples and compare it with live, side by side.
- Per-field control. Formats, value mapping, constraints, lookups, constants and your own instructions.
- A person approves. The system can propose coherence rules. None goes live without approval.
04 · Pipeline
Four phases. One confidence gate.
Each phase fills more of the output grid. The earliest reliable answer wins; a later, less certain guess never replaces it.
- 1 · Resolve. Cells fill straight from the database: no model call, no credits.
- 2 · Agent. Plans the gaps: compute from other values, transfer, re-read or skip with a reason.
- 3 · Validate. Cross-field checks order the review queue. They never block output.
- 4 · Re-read. Empty or weak cells are read again with the field instruction and the whole grid as context.
05 · Rules
Rules a person wrote. Gates a person set.
A rule starts as a sentence in plain words. Talonic compiles it into a check; the sentence is what a reviewer reads when it fires.
- Three effects. Hold for a person, flag and deliver with a mark, or record in the ledger only.
- Honest outcomes. Passed, failed, not applicable or indeterminate. A missing input is never a pass.
- Approval gates. Confidence, pass rate and coverage thresholds per spec. What clears them delivers itself.
- Rules mined from your data. Talonic proposes the rules behind past decisions, with the evidence for and against. You confirm.
06 · Cases
Documents do not live alone.
A shipment is a bill of lading, a rate confirmation and an invoice. Talonic links related documents into cases. The unit of work is the case.
- Linked by shared keys: names, contract and load numbers, project codes and other references.
- No false links. Values found in more than 30% of documents, like a common bank, never form a case.
- Explained. Each case shows its documents, the keys that joined them, a timeline and a summary.
07 · Matching
From reading to reconciliation.
Reading tells you what a document says. Matching tells you what to do with it: against your vendor master, carrier list, chart of accounts or open loads.
- Four strategies, exact, fuzzy, date range and numeric range, weighted into one score.
- Evidence per field. Top five candidates with what each strategy contributed.
- It abstains. When it is not sure, it says so, even across ten thousand open loads.
08 · Delivery
Typed delivery, not a webhook.
Results flow to Dynamics, Ivalua, Salesforce, a warehouse or an agent through one typed pipeline. Every attempt logged, every failure replayable.
- Five stages: signal, binding, resolver, serializer, connector. Misconfigurations fail at setup, not in production.
- Any format: JSON, NDJSON, CSV, XLSX, Markdown and more, with fields renamed or dropped without code.
- Destinations: webhook, SFTP, S3, Azure Blob, Google Drive and Sheets, OneDrive, SharePoint, Gmail, Outlook.
- Safe by default: HMAC-signed, idempotency keys, seven retries over ten hours, a dead-letter queue you can replay.
09 · Query
Stop searching your documents. Query them.
Retrieval re-reads the corpus on every question and can answer differently tomorrow. A database answers once, the same way, with a line number.
- Typed rows. Filter by field value with no model call.
- Cited answers. Ask in plain language. Talonic plans, runs read-only SQL and grounds every claim in a source span.
- Scoped. Limit any question to these documents, this spec or this pipeline, and nothing else.
- Everywhere. In the workspace, through the API, or in Claude and Cursor over MCP.
10 · Agents
An agent is a scope, some rules and a ledger.
Give an agent exactly these contracts and this reference table, nothing else. Every read is logged, every answer carries its source, every decision is recorded.
- You decide what it sees, and you can prove what it saw.
- Rules in plain words decide what it does. People handle what the rules hold.
- Three run on client data today, each with a status in four plain words: running, shadow run, on request, authored.
11 · API
Every primitive has an endpoint.
One typed REST surface for production integrations and agents alike, also shipped as a Node SDK and an MCP server.
- Read: /v1/extract, /v1/documents/filter, /v1/ask.
- Build: /v1/schemas, /v1/structuring/gates, /v1/cases, /v1/matching.
- Ship: /v1/data-products, /v1/delivery, /v1/credits.
- Keys: tlnc_ bearer keys with six scopes, hashed at rest, shown once.
POST /v1/ask
Authorization: Bearer tlnc_...
{
"question": "Which contracts renew in the next 30 days?",
"scope": { "pipeline_id": "pipe_contracts_2026" }
}
→ rows, the SQL behind them, a source span per claim12 · Comparison
What we are. What we are not.
Every choice is a tradeoff. This is what we would tell you in person.
Talonic vs. Reducto →Talonic vs. Instabase →Benchmark against RAG →
| Reducto | Instabase | RAG / copilots | Talonic | |
|---|---|---|---|---|
| Data model | One target schema | One target schema | Chunks and embeddings | A database with provenance |
| Second question, same corpus | Re-parse | Re-parse | Re-retrieve, re-spend | A lookup, zero AI calls |
| Same question tomorrow | Same parse | Same parse | Not guaranteed | Byte-identical |
| Parsing fidelity | Strong | Strong | Variable | Strong |
| Cases across documents | None | None | None | Native |
| Rules, holds and a ledger | None | None | None | Native |
| Typed delivery (HMAC, DLQ, replay) | Partial | Partial | None | Native |
| Published benchmarks, losses included | None | None | None | Three |
| EU data residency by default | Partial | Partial | Variable | Native |
| DIN SPEC 91491 | None | None | None | Co-author |
We are slower out of the gate than parsing-first vendors and not always the cheapest per page. If you want the fastest path from PDF to JSON, Reducto is excellent. If you want the database underneath your document workflows, that is us.
Send us a pile of files. We’ll send back a database.
Send a representative sample. Within five business days you get your data back with field-level coverage, confidence and provenance.
Start the document test30 minutes with an engineer on architecture, throughput, residency and integration. Or start free: 5,000 credits a month, no card.
Read the docsGDPR and HIPAA compliant, ISO 27001 and 42001 aligned, EU-resident infrastructure in Germany, DIN SPEC 91491 co-author.