01 / Extract
Find what matters.
Turn repeated requests for entities, dates, fields or clauses into structured information that stays connected to the source.
AI Initiative / DocuMind / Document intelligence
From documents to understanding you can use.
Explore what your documents contain. Compare interpretations. Turn reviewed information into useful data, with the source and the decision still attached.
A proposed DocuMind journey, developed from the original concept. The selected direction is broad document intelligence, illustrated through contracts. Implementation presentation and architecture choices remain open. This is an explanatory prototype, not a live document-processing service.
One question. An inspectable answer.
DOCUMENTS · ILLUSTRATIVE30 days. Or 60?
A contract and its amendment contain different renewal notice periods.
↓ Find the passages · compare interpretations · review
REVIEWED RECORD · ILLUSTRATIVE60 days
Amendment reference + reviewer decision + output version.
The value is useful. Knowing where it came from makes it accountable.
Information becomes useful when people can check it.
DocuMind explores the space between reading a document and acting on its contents. Contracts make that journey concrete, but the original idea extends to other document types and existing information workflows.
02 / Reconcile
Find what differs.
Compare a document with existing OCR, NLP or manually entered records. Put discrepancies in front of a reviewer.
03 / Explore
Work through the meaning.
Use an AI peer to investigate relevant passages, ask follow-up questions and support a person’s document workflow.
Follow the information.
Select a stage to see what enters, what stays and what moves forward. The record layer preserves context; the return loop turns feedback into evaluated change.
Illustrative journey: a contract says 30 days; an amendment says 60. Follow the evidence through comparison, human review and delivery. No documents are uploaded or processed here.
- Receives
- Retains
- Produces
In the worked example
Review boundary
Read the complete workflow as text
All seven stages and their review boundaries are retained here. This reading view works without JavaScript.
01
LoadStart with the source
Bring a document or a collection into a defined task. Repository connectors, uploads and batch jobs are possible entry points; each needs an explicit access boundary.
- Receives
- Contract v3 and a request for its renewal terms
- Retains
- Source identifier, version, content hash, permission context and intake time
- Produces
- A traceable source record and a processing request
Follow the example and review boundary
An illustrative contract and its amendment are loaded together. The requested field is notice_period_days.
Boundary. Reject unsupported or unreadable inputs visibly. Do not silently omit an attachment.
02
RetainKeep context, not just chunks
Keep the original for the permitted processing and review period. Organise extracted text, page references, sections and metadata so answers can be traced back while the source is available. Derived indexes support discovery; they do not replace source records.
- Receives
- Original documents and processing request
- Retains
- Original version, extracted passages, page/section anchors and derivation links
- Produces
- An organised, searchable document collection
Follow the example and review boundary
Section 8.2 and the amendment retain separate identities. A search result can point to either, with its version.
Boundary. The historical design includes delivery followed by purge. Define retention across originals, extracted text, indexes and cached results; retained metadata cannot recreate deleted evidence.
03
ExploreAsk, search and inspect
A person can explore a collection interactively, or submit a structured set of fields for batch processing. Relevant passages stay alongside the question and proposed interpretation.
- Receives
- A question or schema, with permitted document context
- Retains
- Request, retrieved passage references and retrieval configuration
- Produces
- Relevant evidence and a draft structured answer
Follow the example and review boundary
The question finds a 30-day notice clause and an amendment that changes the period to 60 days.
Boundary. Show missing evidence and conflicting versions instead of producing an unsupported single answer.
04
CompareMake disagreement visible
Separate model passes can extract or critique the same field against source passages. Compare values and supporting evidence, then route disagreements for review. Agreement is a useful signal, not a correctness guarantee.
- Receives
- The same requested field and source-grounded context
- Retains
- Model/version, prompt configuration, candidate value and cited passage for each pass
- Produces
- Agreement, disagreement or insufficient-evidence status
Follow the example and review boundary
Model A proposes 30 days from the main contract; Model B proposes 60 days from the amendment. The difference is highlighted.
Boundary. Shared model biases can produce shared errors. Cross-check against the source and deterministic field constraints.
05
ReviewA person resolves the decision
The reviewer sees competing interpretations and their supporting passages. They can accept, correct, reject or request more evidence, with a reason. Reviewing extracted information is distinct from downstream business approval, which the original concept places outside DocuMind.
- Receives
- Candidate answers, conflicts and source passages
- Retains
- Reviewer action, corrected value, reason, source reference and review time
- Produces
- Reviewed information or an unresolved item
Follow the example and review boundary
The reviewer confirms the amendment applies and records 60 days, with the amendment as evidence.
Boundary. Unresolved items remain unresolved. A model score must not silently substitute for a required human decision.
06
DeliverProduce usable, accountable data
Validate the reviewed record against its output schema and send it to the intended report, table or downstream process. Preserve permitted lineage and the destination acknowledgement. Apply the agreed retain-or-purge policy after delivery; repository management remains an external responsibility unless separately added.
- Receives
- Reviewed fields and a destination-specific schema
- Retains
- Output version, source and review references, validation result and delivery receipt
- Produces
- A structured record, report or reconciliation result
Follow the example and review boundary
The illustrative output is notice_period_days: 60, with source and review references. An existing 30-day record is flagged for correction.
Boundary. Prevent duplicate delivery and make failures recoverable. A correct extraction and a successful destination write are separate outcomes.
07
ImproveLearn through controlled change
Collect corrections and failure patterns for evaluation. First decide whether the issue belongs in parsing, retrieval, instructions or review policy. Model adaptation is a separate option with its own data, training and evaluation requirements.
- Receives
- Reviewed corrections, failures and evaluation cases
- Retains
- Versioned feedback dataset, proposed change, evaluation results and release decision
- Produces
- An evaluated workflow change, or a rejected experiment
Follow the example and review boundary
The amendment conflict becomes a regression case. A revised retrieval or comparison step must find both passages before it is adopted.
Boundary. Recording feedback is not RLHF. Any training experiment needs a defined objective, suitable feedback data and held-out evaluation before rollout.
One workflow. Distinct responsibilities.
This logical architecture separates source custody, AI interpretation, human decisions and delivery. It is a proposed composition for review, not a claim that these services are deployed together.
Return to the connected architecture diagram ↑
Explore technical responsibilities and related systems
People and entry points
Workspace + batch requests
Document selection, task schemas, interactive exploration, comparison views and reviewer actions.
Processing and coordination
Intake → retrieval → interpretation → review → delivery
A workflow coordinator tracks stage state and failures. Document processing preserves anchors; model adapters produce candidates; comparison exposes conflicts; review and output validation control progression.
Persistent foundation / lineage across every step
Separate source custody, interpretations and decisions.
Source versionsPassages + indexesModel candidatesHuman decisionsDelivery recordsEvaluation cases
Shared identifiers connect these records while retention policies govern their lifetime. The original design includes delivery followed by purge; any retained metadata must make the loss of source evidence explicit. Repository management and downstream business approval remain outside the historical scope.
Change loop
Feedback → evaluation → controlled release
Corrections inform a proposed change. Held-out cases test it. An accepted change gets a version and rollback path.
Related architecture / potential reuse
Connect to the wider work deliberately.
Conduir / Role Fabric offers a related coordination architecture; Cortex offers a related memory architecture. These are candidates for reuse, not verified DocuMind integrations. The historical “Powered by Quark” association is retained in the original concept.
Feedback should improve a process you can still inspect.
Human feedback can correct a record today and inform an experiment later. A reviewed correction improves a record now; changing a workflow or model requires a separate evaluation.
| Mechanism | What changes | What would establish it |
|---|---|---|
| Human review | A candidate value becomes accepted, corrected, rejected or unresolved. | A recorded decision with supporting evidence and review identity. |
| Workflow improvement | Parsing, retrieval, instructions, comparison or routing. | A versioned change evaluated against representative and regression cases. |
| Model adaptation, including possible RLHF | Model behaviour through a separately designed training process using suitable feedback. | Training data provenance, a defined objective, training artifacts and held-out evaluation. No such DocuMind evidence has been established in this review. |
From documented intent to a demonstrable implementation.
The original writing gives this page substance. The record boundaries and coordinated lifecycle shown here are proposed extensions for discussion, not recovered implementation specifications.
Recovered
The original product idea
Repository intake, granular extraction, interactive and batch work, model comparison, human correction, reconciliation and downstream delivery are described in the retained writing. The author also records testing with OpenAI and LLaMA2-70B; corresponding test artifacts have not yet been recovered.
Authored here
A coherent end-to-end draft
The seven-stage flow, worked example and explicit record boundaries make the concept reviewable. The example is synthetic and makes no model calls.
Still to establish
Implementation and outcomes
A named DocuMind codebase, working connectors, persisted lineage, comparison behaviour, feedback training and measured quality have not yet been established by this review.
Implementation acceptance path to discuss
- Load a permitted document and preserve versioned passages with source anchors.
- Retrieve and extract a requested field; retain the model candidate and supporting passages.
- Expose a planted contradiction between passages or model outputs.
- Record a reviewer’s correction and prevent an unresolved item from being delivered as accepted.
- Deliver schema-valid data with lineage; test retries and destination failures.
- Turn the correction into an evaluation case and demonstrate whether a proposed change improves results.
Assess extraction accuracy, unsupported answers, disagreement detection, reviewer effort, latency and cost separately. Test permissions and retention across derived records. These are proposed checks, not completed results.
A step towards AI that grows with us.
DocuMind gives the wider vision a practical setting: AI helps people explore information, people remain involved in consequential interpretation, and useful understanding can be carried forward with its context.