"""Screen 2: Dossier folder and profile evidence.
Normal (non-technical) flow — a single button:
* **Continue to Job Screenshot** — the user points the tool at a local
folder and clicks once. That one button runs the full backend chain
silently behind a spinner — ``validate`` → ``read_dossier`` →
``create_evidence_index`` — then advances. The user never sees the
intermediate "Validate Folder" / "Read Dossier" / "Create Evidence Index"
steps, the file count, or the dossier-strength score. A folder that can't
be validated/read shows one clean inline error and keeps the user on the
page — never a traceback.
Original files never leave the user's machine. The file count, the
dossier-strength score, the separate Validate/Read/Index buttons, the
chunk/proof-point stats, per-file source types, and the scoring rubric
appear only when ``SHOW_DEBUG_PANEL=true``.
"""
from __future__ import annotations
import hashlib
import os
import shutil
import tempfile
from contextlib import contextmanager
from pathlib import Path
import streamlit as st
from app.config import clear_opportunity_state, get_settings, save_dossier_path
from app.models.schemas import SOURCE_TYPE_LABELS
from app.services.dossier_reader import (
DossierReadSummary,
read_dossier,
summarize_chunks,
)
from app.services.evidence_index import build_evidence_index
from app.services.folder_validator import (
FolderValidationResult,
SUPPORTED_EXTENSIONS,
THIN_DOSSIER_THRESHOLD,
missing_in_plain_words,
score_evidence,
score_rubric_labels,
validate,
)
from app.ui import interview_screen, theme
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def _strength_color(score: int) -> str:
if score >= 80:
return "green"
if score >= 60:
return "blue"
if score >= 40:
return "orange"
return "red"
def _reset_downstream() -> None:
"""Invalidate everything that depends on the dossier contents."""
for key in (
"dossier_chunks",
"evidence_index",
"canonical_profile",
"evidence_index_meta",
"extracted_job_fields",
):
st.session_state[key] = None
st.session_state.dossier_read = False
# Whatever is built next comes from the folder, not from an earlier
# interview. The interview's own answers are deliberately kept, so a user
# who switches back to it carries on where they left off.
st.session_state["profile_source"] = "dossier"
# The interview publishes a strength score of its own, and until now
# nothing ever cleared it. Left behind, it describes a profile that no
# longer exists: it suppresses the gap-fill offer for a genuinely thin
# folder (see _should_offer_gap_fill) and lets the analysis score this
# folder's profile with the interview's number. The answers stay, so
# resuming the interview publishes it again.
st.session_state["interview_strength"] = None
# Also drop any per-opportunity analysis (match/score/recommendation/
# proposal/fingerprint) — it was built on the old evidence.
clear_opportunity_state(st.session_state)
def _strength_tone(score: int) -> str:
"""The strength label's tone on the canvas's KPI tile.
The thresholds are _strength_color's, unchanged — this only picks which of
the design's bands paints the label.
"""
if score >= 80:
return "ok"
if score >= 60:
return "t2"
if score >= 40:
return "warn"
return "stop"
def _render_validation_clean(result: FolderValidationResult, *, debug: bool) -> None:
if result.issues and not result.exists:
for issue in result.issues:
st.error(issue)
return
# Normal UI shows ONLY the file count. The X/100 dossier-strength score,
# its label, the "to strengthen your dossier consider adding…" line, and
# the "evidence collection looks solid" banner mislead non-technical
# users, so they are gated behind SHOW_DEBUG_PANEL=true.
# Same numbers, same gating — rendered as the design canvas's KPI tiles.
if debug:
theme.kpis(
[
{"lab": "Dossier files", "val": result.total_files, "icon": "folder"},
{
"lab": "Dossier strength",
"val": f"{result.strength_score}/100",
"icon": "spark",
"delta": result.strength_label,
"delta_kind": _strength_tone(result.strength_score),
},
],
columns=2,
)
else:
theme.kpis(
[{"lab": "Dossier files", "val": result.total_files, "icon": "folder"}],
columns=1,
)
for issue in result.issues:
st.error(issue)
for warning in result.warnings:
st.warning(warning)
if debug and result.missing_categories:
st.caption(
"To strengthen your dossier, consider adding: "
+ ", ".join(result.missing_categories)
)
if (
debug
and result.can_continue
and not result.warnings
and not result.issues
and result.strength_score >= 60
):
st.success("Your evidence collection looks solid.")
def _render_debug_panel(result: FolderValidationResult | None) -> None:
settings = get_settings()
meta = st.session_state.get("evidence_index_meta") or {}
proofs = st.session_state.get("evidence_index")
with st.expander("Developer Debug Panel", expanded=False):
st.markdown(
f"**Provider:** `{settings.llm_provider}` • "
f"**Model:** `{settings.active_model}`"
)
chunks = st.session_state.get("dossier_chunks")
if st.session_state.get("dossier_read") and chunks is not None:
summary = summarize_chunks(chunks)
st.markdown(
f"**Dossier read:** {summary.files_processed} file(s), "
f"{summary.chunks_extracted} chunk(s), "
f"{summary.failed_files} failed."
)
for f in summary.files:
label = SOURCE_TYPE_LABELS.get(f.source_type, f.source_type)
line = (
f"- `{f.file_path}` — {label} · status: **{f.status}** · "
f"chunks: **{f.chunk_count}**"
)
if f.warning:
line += f" · ⚠ {f.warning}"
st.write(line)
if proofs is not None and meta:
st.markdown(
f"**Evidence index source:** "
+ ("LLM API" if meta.get("used_api") else "local fallback")
)
if meta.get("error_message"):
st.caption(f"Reason: {meta['error_message']}")
claim_counts: dict[str, int] = {}
for p in proofs:
claim_counts[p.claim_type] = claim_counts.get(p.claim_type, 0) + 1
if claim_counts:
st.markdown("**Proof points by claim type:**")
for claim_type, count in sorted(
claim_counts.items(), key=lambda kv: (-kv[1], kv[0])
):
st.write(f"- {claim_type}: **{count}**")
if result is None:
return
if result.source_type_counts:
st.markdown("**Source-type breakdown:**")
for source_type, count in sorted(
result.source_type_counts.items(), key=lambda kv: (-kv[1], kv[0])
):
label = SOURCE_TYPE_LABELS.get(source_type, source_type)
st.write(f"- {label}: **{count}**")
labels = score_rubric_labels()
if labels:
st.markdown("**Score breakdown:**")
for key, label in labels.items():
awarded = result.score_breakdown.get(key, 0)
st.write(f"- {label}: **{awarded}**")
if result.files:
st.markdown(f"**Files detected ({result.total_files}):**")
for record in result.files:
label = (
SOURCE_TYPE_LABELS.get(record.source_type, record.source_type)
if record.source_type
else "—"
)
icon = (
"✓"
if record.supported and record.readable
else ("⊘" if not record.supported else "✗")
)
st.write(f"{icon} `{record.relative_path}` — {label}")
st.caption(
"Supported extensions: " + ", ".join(sorted(SUPPORTED_EXTENSIONS))
)
# ---------------------------------------------------------------------------
# Cloud (upload) mode
# ---------------------------------------------------------------------------
#
# On a desktop install the user points the app at a local folder. On a shared
# cloud deployment (e.g. a Hugging Face Space) there is no access to the
# visitor's machine, so they upload their files instead. The uploaded bytes are
# held in this session only. They are written to a private temp directory just
# while the existing ``validate → read_dossier → build_evidence_index`` pipeline
# runs against it unchanged, and that directory is deleted as soon as the read
# ends — the server's disk never keeps a visitor's files. Nothing about the
# downstream flow or look changes.
def is_cloud_mode() -> bool:
"""True when running on a shared host where local folder paths can't work.
Detected from Hugging Face's ``SPACE_ID`` (set on every Space) or an
explicit ``UPWORK_CLOUD`` override, so local desktop runs keep the original
folder-path UI untouched.
"""
return bool(os.getenv("SPACE_ID") or os.getenv("UPWORK_CLOUD"))
# Session keys for the persisted upload set. The records (name + bytes) are
# kept in session — NOT just in the file_uploader widget — because the widget
# returns an empty list after the user navigates to another step and back, so
# relying on it alone makes the dossier appear "lost". Kept until the page is
# reloaded (new session) or the user clears them. The widget only ever adds to
# them: whatever is dropped in joins the records and the widget re-mounts
# empty, so an uploader that came back empty is never taken for a smaller set.
# Files leave the set through their own Remove button or Remove all.
_UPLOAD_RECORDS_KEY = "dossier_uploaded_files"
_UPLOAD_SEQ_KEY = "dossier_uploader_seq"
def _upload_signature(records: list[dict]) -> tuple:
"""(name, size, content hash) per file — identifies an upload set."""
return tuple(
(r["name"], len(r["bytes"]), hashlib.sha256(r["bytes"]).hexdigest())
for r in records
)
def _materialize_records(records: list[dict]) -> str:
"""Write the persisted upload records into a new private temp dir.
The directory lives under the OS temp location and is never committed or
shared; the caller deletes it once the read is done (see
:func:`_uploads_on_disk`). Returns the directory path for the standard
folder pipeline.
"""
base_path = Path(tempfile.mkdtemp(prefix="upwork_dossier_"))
taken: set[str] = set()
try:
for r in records:
safe_name = Path(r["name"]).name # strip any directory components
if not safe_name:
continue
# Two uploads can share a name (two "Final.pdf" from different
# folders): the later one is written as "Final (2).pdf" instead of
# overwriting the first. Compared case-blind, as macOS and Windows
# disks compare names.
name, n = safe_name, 1
while name.casefold() in taken:
n += 1
name = f"{Path(safe_name).stem} ({n}){Path(safe_name).suffix}"
taken.add(name.casefold())
(base_path / name).write_bytes(r["bytes"])
except BaseException:
shutil.rmtree(base_path, ignore_errors=True)
raise
return str(base_path)
@contextmanager
def _uploads_on_disk(records: list[dict]):
"""Back ``dossier_folder_path`` with ``records`` for one pipeline run.
The files are written to a new temp dir and the dir is deleted when the
block ends, however it ends — including the ``st.rerun()`` the chain uses
to move on. The bytes stay in session, so the next read simply writes them
out again.
"""
path = _materialize_records(records)
st.session_state["dossier_folder_path"] = path
try:
yield path
finally:
shutil.rmtree(path, ignore_errors=True)
st.session_state["dossier_folder_path"] = ""
def _forget_upload_reading() -> None:
"""The upload set changed: drop everything read from the previous one.
On top of the evidence (:func:`_reset_downstream`), that is the folder
validation — the gap-fill card quotes it, so it would go on describing
files that are gone — the once-only gap-fill stop, so the new set gets
its own offer, and a failed read's error, which would go on asking the
user to replace files already replaced. ``_reset_downstream`` itself
keeps all three, because the Validate and Continue paths call it right
after validating.
A profile the interview built was not read from these files alone,
though. One built from the answers alone ("Interview me instead") never
read them, so it stays exactly as it is. One that added the answers to
them keeps the answers: it falls back to the answers alone, and the next
Continue adds them to whatever files it reads.
"""
st.session_state["dossier_validation"] = None
st.session_state.pop(_THIN_OFFERED_KEY, None)
st.session_state["dossier_read_error"] = False
if st.session_state.get("profile_source") == "interview":
meta = st.session_state.get("evidence_index_meta")
if not (isinstance(meta, dict) and meta.get("with_folder")):
return
if interview_screen.rebuild_profile():
# Still a gap fill, whichever button last opened the interview,
# so the files read next get the answers added back.
st.session_state["interview_entry"] = "gap_fill"
return
_reset_downstream()
def _sync_dossier_uploads() -> list[dict]:
"""Merge whatever was dropped into the uploader into the persisted set.
* Dropped files join the stored records (a file already stored with the
same name and content is not added twice) and invalidate downstream
evidence; the widget then re-mounts empty. Nothing is written to disk
here; the records go to disk only while they are read (see
:func:`_uploads_on_disk`).
* An empty uploader (e.g. after navigating away and back) keeps the stored
records, so the dossier is never silently lost.
Returns the current persisted records.
"""
ss = st.session_state
accepted = sorted(ext.lstrip(".") for ext in SUPPORTED_EXTENSIONS)
uploader_key = f"dossier_uploads_{ss.get(_UPLOAD_SEQ_KEY, 0)}"
files = st.file_uploader(
"Upload your profile evidence",
type=accepted,
accept_multiple_files=True,
key=uploader_key,
help=(
"Your Upwork Profile Optimisation document and your Dossier work "
"best; add a portfolio, résumé or testimonials if you have them. "
"Your files are processed in this session only and are never "
"stored on the server."
),
)
st.caption(
"Upload your **Upwork Profile Optimisation** document and your "
"**Dossier**. Leave out the Offer Bank — it lists offers you didn't "
"choose."
)
stored: list[dict] = ss.get(_UPLOAD_RECORDS_KEY) or []
if files:
known = set(_upload_signature(stored))
added: list[dict] = []
for f in files:
record = {"name": f.name, "bytes": f.getvalue()}
(identity,) = _upload_signature([record])
if identity not in known:
known.add(identity)
added.append(record)
if added:
ss[_UPLOAD_RECORDS_KEY] = stored + added
_forget_upload_reading()
# Empty the widget: its files are in the stored set, listed below it.
ss[_UPLOAD_SEQ_KEY] = ss.get(_UPLOAD_SEQ_KEY, 0) + 1
st.rerun()
return stored
def _remove_dossier_upload(index: int) -> None:
"""Remove one file from the persisted dossier upload set."""
ss = st.session_state
records = list(ss.get(_UPLOAD_RECORDS_KEY) or [])
if 0 <= index < len(records):
del records[index]
ss[_UPLOAD_RECORDS_KEY] = records
_forget_upload_reading()
def _clear_dossier_uploads() -> None:
"""Remove the persisted dossier upload set and reset the uploader widget."""
ss = st.session_state
ss[_UPLOAD_RECORDS_KEY] = []
ss["dossier_folder_path"] = ""
# Bump the uploader key so the widget itself re-mounts empty.
ss[_UPLOAD_SEQ_KEY] = ss.get(_UPLOAD_SEQ_KEY, 0) + 1
_forget_upload_reading()
def _card_upload_dossier(debug: bool) -> None:
"""Cloud variant of the dossier step: upload files, then continue.
Mirrors the single-button normal flow — an uploader plus one *Continue to
Job Screenshot* button that silently runs validate → read → index via
:func:`_continue_chain`. In debug mode the granular read/index cards still
render against the uploaded folder, exactly as on desktop.
"""
with st.container(border=True, key="ups_panel_dossier_1"):
theme.section_label(
"Step 1 · Upload dossier files" if debug else "Upload dossier files"
)
stored = _sync_dossier_uploads()
has_files = bool(stored)
if has_files:
# One row per stored file: the uploader above only adds.
for index, record in enumerate(stored):
col_name, col_remove = st.columns([3, 1])
col_name.text(record["name"])
if col_remove.button(
"Remove",
key=f"dossier_remove_upload_{index}",
use_container_width=True,
help="Remove just this file.",
):
_remove_dossier_upload(index)
st.rerun()
col_ready, col_clear = st.columns([3, 1])
col_ready.caption(f"{len(stored)} file(s) ready.")
if col_clear.button(
"Remove all",
key="dossier_clear_uploads_btn",
use_container_width=True,
help="Clear these files so you can upload a different set.",
):
_clear_dossier_uploads()
st.rerun()
if debug:
# Validate against the uploaded folder so the file count / strength
# panel works the same as the desktop debug flow.
if st.button(
"Validate Files",
key="validate_folder_btn",
disabled=not has_files,
help="Upload at least one file first." if not has_files else None,
):
with _uploads_on_disk(stored) as path:
st.session_state.dossier_validation = validate(path)
_reset_downstream()
result: FolderValidationResult | None = st.session_state.get(
"dossier_validation"
)
if result is not None:
_render_validation_clean(result, debug=debug)
return
if st.button(
"Continue to Job Screenshot",
type="primary",
key="continue_to_screenshot_btn",
disabled=not has_files,
help="Upload at least one file first." if not has_files else None,
):
# Each read gets a fresh temp folder, so the once-only thin-folder
# stop remembers the files themselves rather than their path.
with _uploads_on_disk(stored):
_continue_chain(offer_key=_upload_signature(stored))
elif not has_files:
st.caption(
"Upload your profile evidence files, then continue."
)
# ---------------------------------------------------------------------------
# Cards
# ---------------------------------------------------------------------------
def _card_select_folder(debug: bool) -> None:
with st.container(border=True, key="ups_panel_dossier_2"):
theme.section_label("Step 1 · Select dossier folder")
st.text_input(
"Dossier folder path",
placeholder="/absolute/path/to/your/profile-evidence-folder",
key="dossier_folder_path",
help="Original files never leave your machine.",
)
path = (st.session_state.get("dossier_folder_path") or "").strip()
if st.button(
"Validate Folder",
key="validate_folder_btn",
disabled=not path,
help="Enter a folder path first." if not path else None,
):
st.session_state.dossier_validation = validate(path)
_reset_downstream()
result: FolderValidationResult | None = st.session_state.get(
"dossier_validation"
)
if result is not None:
_render_validation_clean(result, debug=debug)
else:
st.caption("After validation you'll see how many files were found.")
def _resolve_dossier_path(result: FolderValidationResult | None) -> str:
"""Folder path to read: prefer the typed path, fall back to validation."""
path = (st.session_state.get("dossier_folder_path") or "").strip()
if path:
return path
if result is not None and result.exists:
return result.folder
return ""
def _perform_read(path: str) -> None:
"""Read the dossier folder and persist the results into session state.
A single bad file never aborts the read (the reader degrades it to a
``failed`` chunk), and an unexpected top-level failure leaves the user
with a clean error rather than a crashed app.
"""
try:
chunks = read_dossier(path) if path else []
except Exception: # noqa: BLE001 - never crash the screen on a read error
st.session_state.dossier_chunks = []
st.session_state.dossier_read = False
st.session_state.dossier_read_error = True
else:
st.session_state.dossier_chunks = chunks
st.session_state.dossier_read = True
st.session_state.dossier_read_error = False
# A fresh read invalidates anything built from a previous read — and
# anything an interview built: the profile now comes from this folder.
st.session_state.evidence_index = None
st.session_state.canonical_profile = None
st.session_state.evidence_index_meta = None
st.session_state["profile_source"] = "dossier"
st.session_state["interview_strength"] = None
clear_opportunity_state(st.session_state)
def _render_read_summary(summary: DossierReadSummary) -> None:
"""User-facing read result. No raw dossier text — counts only."""
st.success("Dossier read successfully")
col_files, col_chunks, col_failed = st.columns(3)
col_files.metric("Files processed", summary.files_processed)
col_chunks.metric("Chunks extracted", summary.chunks_extracted)
col_failed.metric("Failed files", summary.failed_files)
if summary.failed_files:
st.warning("Some files could not be read, but the app continued.")
st.caption("Profile evidence is held in this session only.")
def _card_read_evidence(cloud: bool = False) -> None:
with st.container(border=True, key="ups_panel_dossier_3"):
theme.section_label("Step 2 · Read profile evidence")
result: FolderValidationResult | None = st.session_state.get(
"dossier_validation"
)
can_read = bool(result and result.can_continue)
if st.button(
"Read Dossier",
key="read_dossier_btn",
disabled=not can_read,
help=None if can_read else "Validate a folder first.",
):
with st.spinner("Reading your profile evidence…"):
if cloud:
# Uploads are on disk only while they are read.
records = st.session_state.get(_UPLOAD_RECORDS_KEY) or []
with _uploads_on_disk(records) as path:
_perform_read(path)
else:
_perform_read(_resolve_dossier_path(result))
# Rendered from session state (not just on click) so the summary
# survives Streamlit reruns.
if st.session_state.get("dossier_read_error"):
st.error(
"We couldn't read that folder. Re-validate the folder and "
"try again."
)
chunks = st.session_state.get("dossier_chunks")
dossier_read = bool(st.session_state.get("dossier_read"))
if dossier_read and chunks is not None:
_render_read_summary(summarize_chunks(chunks))
st.divider()
col_index, _ = st.columns(2)
with col_index:
can_build = bool(dossier_read and chunks)
if st.button(
"Create Evidence Index",
key="build_evidence_btn",
disabled=not can_build,
help=None if can_build else "Read the dossier first.",
):
proofs, profile, meta = build_evidence_index(chunks)
st.session_state.evidence_index = proofs
st.session_state.canonical_profile = profile
st.session_state.evidence_index_meta = meta
_score_read_files(proofs)
proofs = st.session_state.get("evidence_index")
if proofs:
st.metric("Proof Points", len(proofs))
st.caption("Evidence is ready for matching.")
proofs = st.session_state.get("evidence_index")
if proofs:
st.divider()
if st.button(
"Continue to Job Screenshot",
type="primary",
key="continue_to_screenshot_btn",
):
# Persist path so the app auto-reloads it on next restart.
p = (st.session_state.get("dossier_folder_path") or "").strip()
if p:
save_dossier_path(p)
st.session_state.current_step = "screenshot"
st.rerun()
# ---------------------------------------------------------------------------
# Normal-flow continue (runs validate → read → index silently)
# ---------------------------------------------------------------------------
def _score_read_files(proofs: list) -> None:
"""Re-score the folder from what its files turned out to hold.
The validation runs before anything is read and sees file names only, so
a full dossier PDF scored as thin. Scoring the proof points instead is
what decides whether the "less than usual" offer appears.
"""
result = st.session_state.get("dossier_validation")
if result is not None and proofs:
st.session_state.dossier_validation = score_evidence(result, proofs)
def _continue_chain(*, offer_key: object = None) -> None:
"""Run the full dossier pipeline silently, then advance.
Wires the single normal-mode "Continue to Job Screenshot" button to the
backend chain in order — ``validate()`` → ``read_dossier()`` →
``create_evidence_index()`` — behind one spinner, so a non-technical
user never sees the intermediate Validate / Read / Index steps. The
pipeline order is unchanged; only the on-screen surface collapsed to one
action.
A folder that can't be validated (bad path, empty folder, or no readable
files) — or a read that yields nothing — stops the chain: the user stays
on the page and sees one clean error, never a stack trace.
``offer_key`` is what the once-only thin-folder stop remembers: the
folder path unless the caller names something steadier (the cloud
uploader passes its upload signature).
A profile that is the folder plus a gap-filling interview is read back
with the interview in it. The read alone builds the profile from the
files, and the answers would drop out of it without a word.
"""
path = (st.session_state.get("dossier_folder_path") or "").strip()
if offer_key is None:
offer_key = path
# Taken before the reset below clears what they are read from.
was_interview = st.session_state.get("profile_source") == "interview"
keep_interview = interview_screen.adds_to_folder()
with st.spinner("Preparing your profile evidence…"):
# 1. Validate the typed folder. A re-run invalidates any evidence
# built from a previously selected folder.
result = validate(path) if path else None
st.session_state.dossier_validation = result
_reset_downstream()
# 2. Read + 3. build the evidence index only when the folder is usable.
if result is not None and result.can_continue:
_perform_read(path)
chunks = st.session_state.get("dossier_chunks")
if chunks:
proofs, profile, meta = build_evidence_index(chunks)
st.session_state.evidence_index = proofs
st.session_state.canonical_profile = profile
st.session_state.evidence_index_meta = meta
_score_read_files(proofs)
if keep_interview:
interview_screen.rebuild_profile(chunks)
if was_interview and st.session_state.get("profile_source") != "interview":
# The folder has taken the interview's place, so the answers are no
# longer in the profile: a thin folder gets its offer again rather
# than going on as though they were.
st.session_state.pop(_THIN_OFFERED_KEY, None)
if st.session_state.get("evidence_index"):
# Persist the path so the app can auto-reload it on the next restart.
if path:
save_dossier_path(path)
# A thin folder stops here once, because the offer to fill its gaps
# is only drawn on this screen. Pressing Continue again on the same
# folder (or the same uploaded files) goes on, so the offer can never
# become a loop.
if (
interview_screen.is_available()
and _should_offer_gap_fill(result)
and st.session_state.get(_THIN_OFFERED_KEY) != offer_key
):
st.session_state[_THIN_OFFERED_KEY] = offer_key
st.rerun()
st.session_state.current_step = "screenshot"
st.rerun()
else:
st.error(_no_profile_message(result))
# What the cloud screen says when the uploaded files could not be read at all:
# it has an uploader, not a path to check.
_UPLOADS_UNREADABLE = (
"Couldn't read your uploaded files. Remove them and upload them again."
)
# A file the reader opened but took no text from: a scanned PDF's pages come
# back "empty", and an image is kept as metadata only, because nothing here
# runs OCR.
_NO_TEXT_STATUSES = ("empty", "metadata_only")
def _no_profile_message(result: FolderValidationResult | None) -> str:
"""Why Continue built no profile, in words the user can act on.
Only a desktop folder has a path to check. Files that were read but gave
no text — a scanned résumé, a photo of one — are named, because checking
the path or uploading them again would never change them.
"""
cloud = is_cloud_mode()
interview_hint = (
", or answer a few questions above instead"
if interview_screen.is_available()
else ""
)
chunks = st.session_state.get("dossier_chunks") or []
if (
result is None
or not result.can_continue
or st.session_state.get("dossier_read_error")
or not chunks
):
if cloud:
return _UPLOADS_UNREADABLE
return "Couldn't read that folder. Check the path and try again."
# Names go in as code, so Markdown never reads a file name as formatting.
files = summarize_chunks(chunks).files
no_text = [f"`{f.file_name}`" for f in files if f.status in _NO_TEXT_STATUSES]
broken = [f"`{f.file_name}`" for f in files if f.status == "failed"]
where = "" if cloud else " to the folder"
add = "Upload" if cloud else "Add"
if not (no_text or broken):
return (
"We read your files but found no work in them to quote from. "
f"{add} your résumé, portfolio or testimonials{where}"
f"{interview_hint}."
)
message = ""
if no_text:
message += (
f"{_join_words(no_text)} {'has' if len(no_text) == 1 else 'have'} "
"no text the app can read: scanned pages and photos aren't read. "
)
if broken:
message += f"{_join_words(broken)} couldn't be opened. "
return (
message
+ f"{add} a PDF with selectable text, a Word file or a text file{where}"
+ f"{interview_hint}."
)
def _card_continue_normal() -> None:
"""Single-button dossier card for the non-technical flow.
One text box + one "Continue to Job Screenshot" button. The button runs
the whole backend chain silently behind a spinner — ``validate()`` →
``read_dossier()`` → ``create_evidence_index()`` — then advances. There
is no file count, no strength score, and no separate Validate / Read /
Index buttons; those are all debug-only.
"""
# Pre-fill the text input with the persisted path so the user
# sees their last folder already filled in if they land here again.
from app.config import get_saved_dossier_path
if not st.session_state.get("dossier_folder_path"):
saved = get_saved_dossier_path()
if saved:
st.session_state["dossier_folder_path"] = saved
with st.container(border=True, key="ups_panel_dossier_4"):
theme.section_label("Select dossier folder")
st.text_input(
"Dossier folder path",
placeholder="/absolute/path/to/your/profile-evidence-folder",
key="dossier_folder_path",
help="Original files never leave your machine.",
)
path = (st.session_state.get("dossier_folder_path") or "").strip()
# The single button runs validate + read + index, then advances (or,
# on failure, leaves a clean inline error via _continue_chain).
if st.button(
"Continue to Job Screenshot",
type="primary",
key="continue_to_screenshot_btn",
disabled=not path,
help="Enter a folder path first." if not path else None,
):
_continue_chain()
else:
st.caption(
"Point to a local folder of your profile evidence, then continue."
)
# ---------------------------------------------------------------------------
# The two ways into a profile
# ---------------------------------------------------------------------------
#
# The folder route below is untouched: same cards, same single button, same
# backend chain. These two cards only add the second route for users who have
# no files to hand, and for users whose files turned out to be thin.
# The folder (or, on the cloud uploader, the upload set) whose thin score was
# last shown the fill-the-gaps offer.
_THIN_OFFERED_KEY = "dossier_thin_offered"
def _goto_interview(entry: str = "chooser") -> None:
# Both ways in land on the interview's intro card, where the AI toggle
# and "Start over" live. Without this the flag set by an earlier visit
# would drop the user straight back onto the live question with both
# controls out of reach. Answers are kept: with an interview already in
# progress the card's own button reads "Resume the interview".
interview_screen.show_intro_on_entry()
# "chooser" is "Interview me instead": the interview replaces a folder
# read earlier. "gap_fill" adds the interview to the folder.
st.session_state["interview_entry"] = entry
st.session_state.current_step = "interview"
st.rerun()
def _card_choose_path(cloud: bool) -> None:
"""Offer the interview alongside the folder, before the folder cards."""
with st.container(border=True, key="ups_panel_dossier_5"):
theme.section_label("Two ways to build your profile")
col_files, col_talk = st.columns(2)
with col_files:
st.markdown("**Use your own work**")
st.caption(
"Upload your files below."
if cloud
else "Point the app at your folder below."
)
theme.html(theme.tag("Best grounding", "ok"))
with col_talk:
st.markdown("**Answer a few questions**")
st.caption(
"Nothing to upload? The app interviews you and builds the "
"same profile from your answers."
)
if st.button(
"Interview me instead",
key="start_interview_btn",
use_container_width=True,
):
_goto_interview("chooser")
def _card_interview_gap_fill(result: FolderValidationResult, *, debug: bool) -> None:
"""Offer the interview as a gap-filler for a thin dossier."""
with st.container(border=True, key="ups_panel_dossier_6"):
theme.section_label("Your files gave us less than usual")
missing = missing_in_plain_words(result)
st.warning(
"We read your files but couldn't find "
+ (_join_words(missing) if missing else "much to quote from")
+ ". A few questions about your work fill the gaps — your files "
"are kept either way."
)
if debug:
st.caption(
f"Dossier strength {result.strength_score}/100 "
f"(below {THIN_DOSSIER_THRESHOLD})."
)
col_fill, col_go = st.columns([2, 1])
fill = col_fill.button(
"Fill the gaps with a few questions",
type="primary",
key="gap_fill_interview_btn",
use_container_width=True,
)
go_on = col_go.button(
"Continue with my files",
key="continue_thin_folder_btn",
use_container_width=True,
disabled=not st.session_state.get("evidence_index"),
help=(
"Go on to the job post with the profile built from your files."
if st.session_state.get("evidence_index")
else "Read your folder first."
),
)
if fill:
_goto_interview("gap_fill")
if go_on:
st.session_state.current_step = "screenshot"
st.rerun()
def _join_words(items: list[str]) -> str:
"""Join phrases as a sentence would: "a", "a and b", "a, b and c"."""
if len(items) <= 1:
return "".join(items)
return ", ".join(items[:-1]) + " and " + items[-1]
def _should_offer_gap_fill(result: FolderValidationResult | None) -> bool:
"""True when real files were read but they scored thin.
``can_continue`` is what separates a real folder from the placeholder
result the interview leaves behind, so an interview-built profile is never
told its files were disappointing.
The validation is a reading of the folder and nothing else: it keeps its
thin score after a gap-filling interview has already answered for it. So
once the interview has built the profile, or has published a strength of
its own, the offer stops — otherwise "Your files gave us less than usual"
greets the user on every later visit to a gap they have already filled.
"""
if result is None or not result.can_continue:
return False
if st.session_state.get("profile_source") == "interview":
return False
if st.session_state.get("interview_strength") is not None:
return False
return int(getattr(result, "strength_score", 0) or 0) < THIN_DOSSIER_THRESHOLD
# ---------------------------------------------------------------------------
# Public render
# ---------------------------------------------------------------------------
def render() -> None:
if not st.session_state.get("api_ok"):
st.error("This step is locked. Run the API check on the Setup screen first.")
if st.button("Back to Setup", key="back_to_setup_from_dossier"):
st.session_state.current_step = "setup"
st.rerun()
return
settings = get_settings()
debug = bool(getattr(settings, "show_debug_panel", False))
cloud = is_cloud_mode()
intro = (
"Upload your real work — portfolio, résumé, testimonials, past "
"proposals. Every claim is grounded in these files only."
if cloud
else "Point the app at a folder of your real work — portfolio, résumé, "
"testimonials, past proposals. Every claim is grounded in these files only."
)
theme.screen_head("dossier", "Your work dossier", intro)
# If the smart-landing auto-load failed (folder moved / deleted), show a
# clear error so the user knows to update the path — no traceback. The
# cloud screen has no path, only uploads that failed to read.
if st.session_state.get("dossier_read_error") and not st.session_state.get("dossier_read"):
st.error(
_UPLOADS_UNREADABLE
if cloud
else "Your saved dossier folder could not be found or read. "
"Please enter the correct path below and continue."
)
validation: FolderValidationResult | None = st.session_state.get(
"dossier_validation"
)
if interview_screen.is_available():
if _should_offer_gap_fill(validation):
_card_interview_gap_fill(validation, debug=debug)
else:
_card_choose_path(cloud)
if debug:
# Developer/admin mode keeps the granular Validate / Read Dossier /
# Create Evidence Index buttons, the file count, strength score,
# chunk/proof-point stats, and the debug panel. The first card swaps to
# an uploader in cloud mode; everything downstream is identical.
if cloud:
_card_upload_dossier(debug)
else:
_card_select_folder(debug)
_card_read_evidence(cloud)
_render_debug_panel(st.session_state.get("dossier_validation"))
elif cloud:
# Cloud normal mode: one uploader + one button. The button silently
# runs validate → read → index and advances, just like desktop.
_card_upload_dossier(debug)
else:
# Desktop normal mode: one text box + one button. The button silently
# runs validate → read → index and lands the user on the Job Screenshot
# step (no file count, no strength score, no extra buttons).
_card_continue_normal()