Auto-transcribe uploads, one-button generate all, accept-all tags

backend/app/enrichment/describe.py

@@ -33,22 +33,27 @@
 
 Progress = Callable[..., None]
 
+# A description is a retrieval aid, not an essay — but see summarize.py for why this is
+# also told to the model rather than left as a silent backstop: a limit it does not know
+# about is one it writes past, and the truncation lands mid-sentence.
+MAX_DESCRIPTION_CHARS = 6000
+
 SYSTEM_PROMPT = (
     "You write descriptions for a media library. Someone will search this library "
     "later with half-remembered words, and your description is what has to match.\n\n"
-    "Write one paragraph saying what is actually in this item: who appears or speaks, "
-    "what is shown, the places, organisations, events and specific terms a person would "
-    "type. Name things explicitly rather than referring to them generally — 'a man in a "
-    "lab coat' helps nobody find anything; a name does.\n\n"
+    "Write what is actually in this item: who appears or speaks, what is shown, the "
+    "places, organisations, events and specific terms a person would type. Name things "
+    "explicitly rather than referring to them generally — 'a man in a lab coat' helps "
+    "nobody find anything; a name does.\n\n"
+    f"Keep the whole reply under {MAX_DESCRIPTION_CHARS} characters — a hard limit, not "
+    "a target, so finish the sentence you are on rather than trailing off as you "
+    "approach it. Most descriptions need nowhere near this much; write one paragraph "
+    "unless the material genuinely needs more than one.\n\n"
     "Describe, do not judge. Do not assess whether anything shown or said is true, and "
     "do not summarise the argument — another field does that. Never open with 'This "
-    "image' or 'The video shows'. No heading, no preamble, nothing after the paragraph."
+    "image' or 'The video shows'. No heading, no preamble, nothing after the text."
 )
 
-# A description is a retrieval aid, not an essay. The model is asked for one paragraph;
-# this is the backstop for when it does not listen.
-MAX_DESCRIPTION_CHARS = 2000
-
 
 def _prompt(asset: Asset, material: source.SourceMaterial) -> str:
     parts = [f"Filename: {asset.original_name or asset.name}"]

backend/app/enrichment/generate_all.py

@@ -0,0 +1,55 @@
+"""The generate-all job: summarise, describe and tag one asset in a single press.
+
+Three separate provider calls, run one after another — not the merged, single-prompt
+version `docs/m6-ai-enrichment.md` sketches as a future cost optimisation ("one pass,
+not three"). That would mean redesigning all three prompts around one shared
+completion; this is the "press each button in turn" version turned into one job, which
+is what a single "Generate all" control needs and costs nothing extra to get right.
+
+Order is summarize, describe, autotag — not the order the buttons sit in the UI.
+`describe`'s own prompt already knows to skip repeating an existing summary ("A summary
+already exists; do not repeat it"), so running summarize first is what lets that check
+actually have something to check against. `autotag` reads both fields as context, so it
+goes last regardless.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Callable
+
+from sqlmodel import Session
+
+from app.enrichment import autotag, describe, summarize
+from app.models.asset import Asset
+
+logger = logging.getLogger(__name__)
+
+Progress = Callable[..., None]
+
+_STEPS = (
+    ("Summarising", summarize.run),
+    ("Describing", describe.run),
+    ("Suggesting tags", autotag.run),
+)
+
+
+def run(session: Session, asset: Asset, progress: Progress) -> str:
+    """Run summarize, describe and autotag over one asset, in turn.
+
+    Stops at whichever step first raises — the same "one clear error" behaviour a
+    single-action job already has, rather than swallowing a failure to force the
+    remaining steps to run against material that step already showed is unusable.
+    """
+    results = []
+    total = len(_STEPS)
+    for index, (stage, step) in enumerate(_STEPS):
+        base = int(index * 100 / total)
+
+        def inner(_stage: str, pct: int, detail: str = "", *, base=base) -> None:
+            progress(stage, base + pct // total, detail)
+
+        progress(stage, base, "")
+        results.append(step(session, asset, inner))
+
+    return " · ".join(results)

backend/app/enrichment/summarize.py

@@ -26,22 +26,28 @@
 
 Progress = Callable[..., None]
 
+# Enough for several real paragraphs and not enough for an essay. Also told to the
+# model in the prompt below, which is what stops the backstop below from being the
+# thing that actually decides where a summary ends — a limit a model does not know
+# about is a limit it will write past, and the truncation lands mid-sentence.
+MAX_SUMMARY_CHARS = 6000
+
 SYSTEM_PROMPT = (
     "You write summaries for a media library. The person reading yours is trying to "
     "find this file again later, sometimes years afterwards, often remembering only "
     "roughly what was in it.\n\n"
-    "Write one paragraph of plain prose. Lead with what the thing actually is, then "
-    "what it covers — the specific names, places, claims and terms someone would "
-    "search for. Prefer the concrete over the general.\n\n"
+    "Write plain prose. Lead with what the thing actually is, then what it covers — "
+    "the specific names, places, claims and terms someone would search for. Prefer "
+    "the concrete over the general.\n\n"
+    f"Keep the whole reply under {MAX_SUMMARY_CHARS} characters — a hard limit, not a "
+    "target, so finish the paragraph you are on rather than trailing off mid-sentence "
+    "as you approach it. Most summaries need nowhere near this much; write one "
+    "paragraph unless the material genuinely needs more than one.\n\n"
     "Never open with a phrase like 'This video' or 'The transcript shows'. Do not "
     "editorialise, do not assess whether anything said is true, and do not add a "
-    "preamble, a heading, or anything after the paragraph."
+    "preamble, a heading, or anything after the text."
 )
 
-# Enough for a real paragraph and not enough for an essay. The model is told to write
-# one paragraph; this is the backstop for when it does not listen.
-MAX_SUMMARY_CHARS = 2000
-
 
 def _prompt(asset: Asset, material: source.SourceMaterial) -> str:
     parts = [f"Filename: {asset.original_name or asset.name}"]

backend/app/jobs/enrichment.py

@@ -24,6 +24,7 @@
 from app.enrichment.describe import run as run_describe
 from app.enrichment.extract_text import TextExtractionError
 from app.enrichment.extract_text import run as run_extract_text
+from app.enrichment.generate_all import run as run_generate_all
 from app.enrichment.source import NoSourceMaterial
 from app.enrichment.summarize import run as run_summarize
 from app.enrichment.transcribe import TranscriptionError
@@ -45,6 +46,7 @@
     KIND_DESCRIBE,
     KIND_EMBED,
     KIND_EXTRACT_TEXT,
+    KIND_GENERATE_ALL,
     KIND_SUMMARIZE,
     KIND_TRANSCRIBE,
     LIBRARY_KINDS,
@@ -230,6 +232,8 @@ def _run_job(job_id: str) -> None:
                 detail = run_autotag(session, asset, progress)
             elif job.kind == KIND_DESCRIBE:
                 detail = run_describe(session, asset, progress)
+            elif job.kind == KIND_GENERATE_ALL:
+                detail = run_generate_all(session, asset, progress)
             elif job.kind == KIND_BULK_ENRICH:
                 bulk = run_bulk(session, job.user_id, job.payload, progress)
                 detail = f"{bulk.done} done"

backend/app/models/job.py

@@ -31,6 +31,10 @@
 # the odd one out of the per-asset set: it calls no provider, costs nothing, and its
 # output is an input to the other three rather than something a user reads directly.
 KIND_EXTRACT_TEXT = "extract_text"
+# The "Generate all" button: summarize, describe and autotag run one after another as a
+# single job, rather than three activity rows the user has to watch separately. See
+# `enrichment/generate_all.py` for why that order and not the button order.
+KIND_GENERATE_ALL = "generate_all"
 # One action applied across a chosen set of assets. Which action, and which assets,
 # live in `EnrichmentJob.payload` — see there for why it is one job and not N.
 KIND_BULK_ENRICH = "bulk_enrich"
@@ -44,6 +48,7 @@
         KIND_AUTOTAG,
         KIND_EMBED,
         KIND_EXTRACT_TEXT,
+        KIND_GENERATE_ALL,
     }
 )
 

backend/app/routers/enrichment.py

@@ -30,6 +30,7 @@
     KIND_DESCRIBE,
     KIND_EMBED,
     KIND_EXTRACT_TEXT,
+    KIND_GENERATE_ALL,
     KIND_SUMMARIZE,
 )
 from app.models.document import DocumentPage
@@ -182,6 +183,41 @@ def start_describe(
     return DataResponse(data=KINDS["enrichment"].to_activity(job))
 
 
+@router.post(
+    "/{asset_id}/generate-all", response_model=DataResponse[ActivityJobRead], status_code=202
+)
+def start_generate_all(
+    asset_id: str,
+    user: CurrentUser,
+    session: Session = Depends(get_session),
+) -> DataResponse[ActivityJobRead]:
+    """Queue summarize, describe and autotag together — the one-button version of
+    pressing each in turn."""
+    asset = _owned_asset(asset_id, user.id, session)
+    _require_provider(session, user.id)
+
+    if not summarisable(asset):
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail={
+                "code": "not_summarisable",
+                "message": "This asset has no content to generate from",
+            },
+        )
+
+    if enrichment_jobs.active_job(session, asset.id, KIND_GENERATE_ALL) is not None:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail={
+                "code": "already_running",
+                "message": "This asset is already generating metadata",
+            },
+        )
+
+    job = enrichment_jobs.submit(session, asset, KIND_GENERATE_ALL)
+    return DataResponse(data=KINDS["enrichment"].to_activity(job))
+
+
 class DocumentPageRead(BaseModel):
     model_config = ConfigDict(from_attributes=True)
 

backend/app/services/assets.py

@@ -89,6 +89,7 @@ async def ingest_upload(
 
     _reindex(session, asset)
     _describe(session, storage, asset)
+    _chain_transcription(session, asset)
     return asset
 
 
@@ -141,6 +142,47 @@ def _describe(session: Session, storage: LocalStorage, asset: Asset) -> None:
     _reindex(session, asset)
 
 
+def _chain_transcription(session: Session, asset: Asset) -> None:
+    """Transcribe audio and video the moment they land, without being asked.
+
+    Mirrors `jobs/enrichment.py::_chain_embedding`, one step earlier in the same
+    chain: a video that nobody transcribes has nothing for describe, summarise,
+    autotag or search to read, and asking a user to press Transcribe by hand before
+    any of that works is a step the pipeline does not need them for.
+
+    Guarded on a Deepgram key actually being configured, for the same reason the
+    embedding chain is guarded on a provider: queueing unconditionally would put a
+    red "no Deepgram key" row in the activity feed after every single upload, which
+    trains people to ignore it.
+
+    Imports `app.jobs.enrichment` lazily rather than at module load — that module's
+    import chain leads back through `app.enrichment.describe` to this one, and
+    importing it at the top of this file would be a cycle.
+    """
+    from app.enrichment.transcribe import can_transcribe
+    from app.jobs import enrichment as enrichment_jobs
+    from app.models.job import KIND_TRANSCRIBE
+    from app.settings_store import load_deepgram_key
+
+    if not can_transcribe(asset):
+        return
+
+    try:
+        if not load_deepgram_key(session, asset.user_id):
+            logger.info(
+                "Not transcribing asset %s: no Deepgram key configured for this user",
+                asset.id,
+            )
+            return
+
+        if enrichment_jobs.active_job(session, asset.id, KIND_TRANSCRIBE) is not None:
+            return
+
+        enrichment_jobs.submit(session, asset, KIND_TRANSCRIBE)
+    except Exception:  # noqa: BLE001 - an upload that succeeded must stay succeeded
+        logger.warning("Could not queue transcription for asset %s", asset.id, exc_info=True)
+
+
 def _reindex(session: Session, asset: Asset) -> None:
     """Keep the keyword index in step with the row.