Extraction is configured for your organization, not a generic default.
Saberra sets three extraction parameters per client at deployment: the language records are written in, how densely the extraction operates, and how aggressively language is normalized across the record set. These parameters are part of the deployment configuration for each organization and are not shared between clients.
Each parameter is configured as an environment variable on the client's Sera Worker instance. Changes take effect at the next extraction cycle. Parameters are set as part of initial deployment and reviewed during onboarding; they can be adjusted at any point by the Saberra deployment team.
Extraction language
EXTRACTION_LANGUAGE| Value | What it does |
|---|---|
| Englishdefault | All extracted field values written in English regardless of source language. |
| Spanish | All extracted field values written in Spanish. |
| Dutch | All extracted field values written in Dutch. |
The extraction language controls the language of structured record fields (decision text, task descriptions, risk notes, etc.), not the language of source documents. A Spanish-language meeting can be extracted into English records.
Extraction granularity
EXTRACTION_GRANULARITY| Value | What it does |
|---|---|
| essential | Only the most significant decisions, commitments, and risks. Minimal record volume. |
| standarddefault | Balanced extraction. The default for most organizations. Decisions, tasks, risks, roles, and policy notes at normal density. |
| full | Maximum extraction density. Every decision candidate, task, weak signal, and policy note surfaced for reviewer consideration. |
Granularity does not change what qualifies for human review. It changes how many candidates are proposed for the reviewer to accept, edit, or reject. Higher granularity means more review queue volume.
Record correction mode
RECORD_CORRECTION_MODE| Value | What it does |
|---|---|
| A | No language normalization. Field values are written exactly as extracted. |
| B | Light normalization. Obvious inconsistencies (capitalization, punctuation) corrected. |
| Cdefault | Standard normalization. Terminology, role names, and organization names aligned to established usage within the record set. |
| D | Full normalization. Active correction of naming drift, role title inconsistencies, and canonical term enforcement. |
Mode D is most useful for organizations with significant naming variation across meetings and email threads (role titles used inconsistently, different abbreviations for the same team or project). Mode A is appropriate when preserving exact source language is required.
Additional per-client context
Beyond these three parameters, each client's Sera Worker runs with an EXTRACTION_ADDENDUM , a short block of additional instructions appended to the extraction system prompt. This is used to communicate organization-specific context that is not captured by the structured parameters above: organizational type, sector, terminology conventions, or scope boundaries. The addendum is written by the Saberra deployment team during onboarding.
What adaptation does not change
- The human review requirement. Every extraction candidate, regardless of granularity or language, goes through the review queue before entering memory.
- Confidentiality enforcement. Restricted records are excluded from AI retrieval at the data layer, independent of extraction configuration.
- Source citations. Every approved record carries a link to the source meeting or email it came from.
- The record schema. The 26-database structure is the same across all clients. Adaptation changes what is extracted and how, not what the system records or how it stores it.
- Data isolation. Each client's extraction pipeline runs against only that client's sources. Configuration is per-client and parameters are not shared.