Fetch transcript via Supadata, chunk, embed. Raises HTTPException on failure.

Return the module-level Supadata singleton. The client is created once on first call and reused for all subsequent calls. This mirrors t

Enrich raw repository chunks (id, video_id, content, ts...) with video title/url and reshape to the canonical citation-chunk dict. Fetches vi

Return a 'miss' result dict for a case that could not be evaluated.

Mint a JWT and attach it to the response as an httpOnly session cookie.

Split *text* into pieces each ≤ max_chars. Tries separators in order: double-newline, single-newline, sentence, then hard-cut.

Append ``text`` to the buffer; return what is safe to emit now.

Return any held text at end-of-stream (markers stripped).

Return True iff *email* matches the configured admin (case-insensitive). Returns False when ADMIN_USER_EMAIL is unset — fail-safe so a missing co

Run migrations in async mode using the injected async engine.

Write current metrics as the new baseline.

Run seed only if needed: - If SEED_ENABLE is false: skip unconditionally (prod default). - If no videos: full seed (create videos + chunk

Stable signed 64-bit key for `pg_advisory_xact_lock` derived from the UUID. `pg_advisory_xact_lock(int8)` wants a bigint; we XOR the two 64-bit h

Build a DoclingDocument from a video title and transcript text. Uses TITLE, SECTION_HEADER, and PARAGRAPH labels as specified.

Build a minimal DoclingDocument from a single text string (used for segments).

Run chunking + embedding for a video that already exists in the DB. Used when videos were seeded without chunks (e.g., from Sprint 1). Re

Invalidate the session cookie on the client. Browsers only honour a `Set-Cookie` deletion directive when every attribute matches the original

Return the real client IP. With uvicorn's `--proxy-headers --forwarded-allow-ips="*"`, Starlette has already rewritten `request.client` to th

Collapse multiple chunks from the same video into a single citation entry. After the ``is_cited`` pass, chunks from the same video are redundant

Split any chunk that exceeds *max_chars* characters into smaller pieces. Tries to split at paragraph boundaries first, then at sentence bound

Use ffmpeg to make a 16 kHz mono Ogg-Opus file. Whisper's accepted formats are flac/m4a/mp3/mp4/mpeg/mpga/oga/ogg/wav/webm, and rejects bare

Normalize the search response shape. Circle's `/community_members/search` historically returned the member object directly. Newer responses s

Scrape og:description from the YouTube video page as a fallback. YouTube embeds the video description in: <meta property="og:description"

Search parent directories for .env and load it. In containerized deploys there is no .env file on disk — env vars are injected by docker-comp

When HybridChunker returns fewer than 2 chunks for ≥3 paragraphs, split the paragraphs into at least 2 groups manually. Each group is pre

Whisper segments → markdown body with `## [HH:MM:SS]` headings.

Build the YAML frontmatter block that ingest_dynamous_content reads.

Return the module-level Supadata singleton.

Greedily group *parts* (joined by *sep*) such that each group is ≤ max_chars. Parts that are individually larger than max_chars are recur

Process one transcript file. Idempotent.

Full ingest pipeline for a single video: 1. Create video record 2. Chunk transcript 3. Embed chunks (batched) 4. Store ch

Return a fresh AsyncClient bound to the FastAPI app.

If the conversation title is still the default, auto-generate one from the first user message (simple truncation for Sprint 2; LLM-based in S

Build a SimpleNamespace that mimics the SDK's Transcript object with list content.

Pick a stable output path under output_dir based on the lesson_map row.

Walk the root folder by name components to find the file ID for drive_path. drive_path is a forward-slash separated path under the Drive root, e.

Run an SSE token chunk through ``stripper``; return a rewritten chunk or ``None`` if the entire token content was held back. Non-token chunks

Call Whisper with verbose_json so we get segment timestamps. Returns the parsed response dict (includes 'segments' list).

Send a user message and stream the RAG-grounded assistant response. Returns: StreamingResponse with Content-Type: text/event-stream

Verify signature + expiry and return the decoded payload. Raises TokenError on failure.

Embed a single text string via OpenRouter. Args: text: A non-empty string to embed. Returns: A list of 1536 floats repr

* Normalize FastAPI's polymorphic `detail` field into a human-readable string. * * Handlers return `{"detail": "..."}` for most errors, but the defa

Resolve DATABASE_URL, preferring the environment over the ini value. alembic.ini ships with `sqlalchemy.url = env(DATABASE_URL)`, but Alembic

Hash a plaintext password with bcrypt. Returns UTF-8 string for DB storage.

Walk *content_dir* and upsert all `*.md` transcripts. Returns a counts dict: `{scanned, unchanged, ingested, errors}`. Idempotent on no-cont

Insert all 10 seed videos into the database with full chunking + embedding.

Compose the live status line. `label` is the tool-aware text from the * backend (issue #223); `subject` is the legacy raw query/id kept as a * fal