qfa.services.coding_trends#

Deterministic, non-LLM coding-trend table.

Counts coding labels over time periods, assembled from feedback-record metadata. Best-effort: when created is absent or unparseable the table is omitted (None) and the reduce step degrades to text-only synthesis. No LLM, no port — pure services logic.

The period granularity is configurable (day / week / month) so a one-month corpus can still show meaningful trend buckets. week uses ISO week numbering (YYYY-Www) so the ISO year — not the calendar year — anchors the bucket, avoiding the silent off-by-one where 2024-12-30 would otherwise collide with truly-January-2024 records.

Functions

`build_coding_trend_table`(records, *, code_fields)	Build a code-by-period count table from record metadata.
`render_coding_trend_table`(table)	Render the table as a compact text grid for the reduce prompt.

qfa.services.coding_trends.build_coding_trend_table(records: tuple[FeedbackRecordModel, ...], *, code_fields: Sequence[str], period: Literal['day', 'week', 'month'] = 'week') → CodingTrendTable | None[source]#

Build a code-by-period count table from record metadata.

Parameters:

records (tuple[FeedbackRecordModel, ...]) – The full input record set.
code_fields (Sequence[str]) – Metadata keys holding coding labels (comma-separated strings).
period (TrendPeriod) – Bucket granularity. week (the default) is usually right; month is better for multi-year corpora; day for short-window deep-dives.

Returns:

The assembled table, or None when no record carries a parseable date in created (best-effort omission).

Return type:

CodingTrendTable | None

qfa.services.coding_trends.render_coding_trend_table(table: CodingTrendTable) → str[source]#

Render the table as a compact text grid for the reduce prompt.

Rows are codes, columns are periods, cells are integer counts. This is the faithfulness anchor the synthesis prompt cites.