Document intake
Receive documents and preserve their identity and structure.
- Receives
- A source document and permitted use
- Produces
- Identifiable document material
AI Initiative · early concept
DocuMind: the original concept architecture
An explanation of the proposed document-to-reviewed-information flow, with source support and human review at its centre.
Historical proposal A present-day explanation of the original concept, not a claim of a currently available product.
The design question
How could document assistance remain connected to its sources?
The earlier concept explores using language models to help people find and interpret relevant information in documents. This reconstruction explains the responsibilities such a workflow would need; it does not turn the original proposal into a verified implementation.
The useful distinction is between finding a passage, proposing an interpretation and accepting information for a particular use. Keeping those steps inspectable would let a reader see both the source and the remaining uncertainty.
A view of the responsibilities
From documents to reviewed information
- 01Receive and organiseA proposed responsibility
- 02Ask a bounded questionA proposed responsibility
- 03Propose with source supportA proposed responsibility
- 04Review and correctA proposed responsibility
- 05Use the reviewed informationA proposed responsibility
Across the flow
Keep sources, permission, uncertainty and the person’s ability to correct the result visible throughout.
What each part is responsible for
The building blocks
Content organisation
Arrange passages and metadata for subsequent inspection.
- Receives
- Document content and source structure
- Produces
- Organised passages with their source context
Retrieval and extraction
Find relevant passages and propose structured information.
- Receives
- A question or requested field and relevant passages
- Produces
- A proposed answer tied to source material
Comparison and human review
Examine disagreements, omissions and source support.
- Receives
- Proposed information and the original passages
- Produces
- A checked result or a recorded uncertainty
Reviewed output
Provide information for an explicitly defined downstream use.
- Receives
- Reviewed information and its intended use
- Produces
- A bounded output suitable for the chosen workflow
Feedback and further evaluation
Record corrections; any model training would require a separately designed and evaluated process.
- Receives
- Corrections and review observations
- Produces
- A record for later evaluation
Following one piece of work
How the flow fits together
01Receive and organise
Keep source identity and the structure needed to interpret each passage. A date or obligation can mean something different when removed from its surrounding clause; text extraction alone does not resolve that meaning.
02Ask a bounded question
Make the requested information clear and search within material appropriate to that request. The concept should allow the system to say that supporting material is missing rather than fill an answer with plausible wording.
03Propose with source support
Provide a candidate result with the passage that supports it. Any structured extraction should remain connected to the source so the reader can inspect ambiguity, exceptions and missing context.
04Review and correct
Let a person compare the result with the original document and record corrections or uncertainty. Several models producing the same answer does not replace this examination, and a correction log is not automatic model training.
05Use the reviewed information
Pass the checked information into its intended workflow while retaining its source context and limits. Any later learning or deletion process would need its own design, evaluation and account of derived information.
Illustrative example · not a benchmark result
Find a renewal clause
A reviewer asks where a contract states a renewal date. A proposed result points to the relevant passage so the reviewer can confirm its meaning.
- The proposed flow would locate relevant source material, identify ambiguity and return a result for inspection.
- If a source is missing or several interpretations remain possible, the result should preserve that uncertainty for the reviewer.
The useful result: Information that can be checked against its source before use.
This is an illustration of the concept, not evidence that the historical product was implemented or that its original guarantees were achieved.
Reading the architecture accurately
Scope and status
What the original concept describes
- Receive documents and preserve their identity and structure.
- Arrange passages and metadata for subsequent inspection.
- Find relevant passages and propose structured information.
What the concept does not establish
- The retained writing is a concept proposal, not a current product or accuracy certificate.
- Several models agreeing does not guarantee that the extracted information is correct.
- Correction records do not automatically retrain a model.
- Removing a source file does not by itself establish that derived records or learned information have been forgotten.
- Legal interpretation requires appropriate human judgement and review of the original document.
Where the design must hold up
Boundaries and evaluation
| Boundary | Why it matters | How to examine it |
|---|---|---|
| Source-supported interpretation | Several models agreeing does not guarantee that extracted information is correct. | Check the proposed result against the original passage, including exceptions, surrounding context and unresolved ambiguity. |
| Correction versus training | A correction record does not automatically retrain a model. | Identify whether feedback is stored, used to revise a result or included in a separately designed and evaluated learning process. |
| Retention and derived records | Removing a source file does not by itself show that its derived records or learned information have been forgotten. | Follow the intended retention or removal operation across the relevant derived material before making a forgetting claim. |
| Historical availability | The retained writing is a concept proposal, not a current product or accuracy certificate. | Establish implementation, evaluation and operating behaviour separately before making a present-day capability claim. |