Portfolio Ingest

Portfolio Ingest

Proven end-to-end on Cortland Batch 1 (2026-07-02): 1 org, 10 assets, 10 Audette links (5 required property grouping), 9/10 ESPM links, 30/30 documents indexed (7,406 chunks). Follow the stages in order — ordering constraints are real.

Relationship to the frontend onboarding wizard

The platform UI has an interactive flow (AssetOnboardingModal → AssetReviewTable → AssetCreationRunner in platform-web) that does drag-drop → LLM extraction → geocode/footprint → fuzzy-match review → create+link+upload. This skill and the wizard call the same soapbox-api endpoints and must stay convention-compatible:

• Same endpoints: POST /api/assets, PATCH /api/assets/:id, GET/POST /api/portfolios/audette|espm/properties[/refresh], POST /api/assets/:id/files, POST /api/portfolios/files.
• Same auto-match threshold: ≥ 0.85 auto-link, 0.40–0.85 review, < 0.40 no match.
• Same metadata convention: metadata.setup_complete: false + metadata.ingestion_source — the wizard writes "asset-onboarding", this skill writes "portfolio-onboarding". Both are recognized downstream; keep the distinction so runs are attributable.
• Division of labor: the wizard assumes the org/portfolio already exist and works asset-by-asset with a human reviewing matches. This skill additionally creates the org/portfolio/members, does Audette property grouping (the wizard can only link or open a "Create in Audette" thread), handles portfolio-level reference files, splits oversized PDFs, retries failed indexing, and enriches asset metadata from helper data. If the user is present and wants to eyeball matches, you can stop after Stage 2 and hand them to the UI ("Add assets") instead.
• Footprint mapping ("📍 Map" chip — Overture footprint queue) is the wizard's specialty; skip it here unless asked.
• BUT backfill the postal-address fields (street_address, city, state_province, postal_zip_code, country) when you link Audette. Audette's building model (get_building_model_details) carries the full address for every property, and you are already pulling that model to link the asset — copy the address into those columns in the same asset-update call. Do NOT leave them null (the Greystar ingest did: 37/39 had lat/lon from geocoding but 0/39 had street/city/state, so every downstream run — BPS applicability, CRREM region, RUBS/VNM legislation, physrisk, brave lookups — had to re-pull addresses from Audette). state_province especially gates the jurisdiction checks. If Audette is unavailable, reverse-geocode from lat/lon (Overture) as a fallback and label (est.).

Prerequisites

• Service account: Vaultwarden item Soapbox Service Account (claude@agents.soapbox.build) (vw get password "..."). Sign in via Supabase auth REST (POST {SUPABASE_URL}/auth/v1/token?grant_type=password with the anon key from ~/platform-web/.env.local). Tokens last 1h — re-auth before long stages.
• API: https://soapbox-api-production.up.railway.app; every call needs Authorization: Bearer <token> + x-organization-id: <org uuid> (tenant middleware resolves the portfolio from the org).
• Supabase project fplbvanvwvnviczozwhz (direct SQL via Supabase MCP) for the steps with no API endpoint: org + member creation, connector-row copies, metadata jsonb merges, and verification rollups. All asset fields go through the API.
• Scripts (already built, scripts/): portfolio_match.py (fuzzy matching), document_classifier.py (doc typing), ll_allocation.py (LL/TT tree). Use them when name matches aren't obvious; skip when filenames are unambiguous.

Stage 0 — Source discovery

Accept any of: zip/Taildrop drop, Google Drive/Box folder, typed register. Extract zips excluding __MACOSX/ and .DS_Store. Dedupe by sha256 — client drops routinely duplicate files across folders (Cortland: 31 of 60 were dupes). Prefer the copy inside a typed subfolder (PCAs/, Audits/) as canonical.

Build the asset register: {name, address, city, state, zip, gfa, year_built, floors, units} per asset from helper data / spreadsheet / audits. Preserve any confidence tags — mark everything not client-verified as such in metadata.

Stage 1 — Org + portfolio + members

1. SQL: insert into organizations (name) values ('<Client>'); add organization_members rows for christopher@soapbox.build (f243660a-6991-4f7a-97f4-fefce9e24873) and claude@agents.soapbox.build (63025178-e0ac-4018-910d-55b876e441db), role owner.
2. API: POST /api/portfolios {name, organization_id} as the service account — this provisions the Stripe customer, core plugins (plugin_catalog auto- provision), and the caller's portfolio membership. Do NOT insert the portfolio by SQL.
3. SQL: add christopher as portfolio_members admin (the API only adds the caller).

Stage 2 — Create assets (BEFORE installing the Audette connector)

Ordering constraint: POST /api/assets fires triggerAudettePropertyCreate, which auto-creates Audette properties if plugin_audette is installed on the portfolio. Creating assets first (no connector yet) makes the pipeline skip (audette_pipeline_state: skipped_no_plugin) and prevents duplicate Audette properties when records already exist.

Per asset:

• POST /api/assets {name, lat?, lon?, num_floors?}.
• PATCH /api/assets/:id {metadata} — include ingestion_source: "portfolio-onboarding", batch, setup_complete: false, num_buildings, archetype, regulatory_driver, energy (with source + units), equipment_survey, and confidence notes.
• PATCH /api/assets/:id also accepts (since soapbox-api 962ab29, 2026-07-02): street_address, city, state_province, postal_zip_code, country, property_type, gross_floor_area_m2 (ft² × 0.09290304), year_built, num_floors, tags (tag with portfolio-onboarding + <client>-batch-N). No direct SQL needed for asset fields — everything goes through the API.

Stage 3 — Audette linking + grouping

1. Install the connector: copy the freshest plugin_audette row in asset_connectors (any portfolio; tokens are user-scoped and work across portfolios; the connector proxy refreshes them) into a new row for this portfolio.
2. GET /api/portfolios/audette/accounts → find the client account uid → PATCH /api/portfolios {audette_account_id} (clears the properties cache).
3. POST /api/portfolios/audette/properties/refresh → per-building rows with property_uid, property_name, building_uid, building_name.
4. Match each asset by name (exact/fuzzy; portfolio_match.py for ambiguous):
   • Grouped (has property_uid) → PATCH /api/assets/:id {audette_property_id}.
   • Ungrouped (buildings match by name but property_uid is null — common; half of Cortland): collect the building uids, then call the Audette MCP directly (https://mcp-server.prod.audette.io/mcp, bearer = connector api_key; JSON-RPC tools/call): switch_customer_account {customer_account_uid} then create_property_for_building {building_model_uids: [...], property_name} → returns property_uid → PATCH the asset. Also assign_property_to_building exists for adding buildings to an existing property. Store the building uids in metadata.audette_building_uids regardless (analysis tools take building_model_uid).
   • No match → leave null; the asset card shows no Audette badge; record as follow-up.
5. Re-run .../audette/properties/refresh at the end so the cache reflects new groupings.

Stage 4 — ESPM linking (always optional, never blocking)

1. Copy a plugin_energy_star connector row (shared ESPM account) to the portfolio.
2. POST /api/portfolios/espm/properties/refresh (~2,300 properties; cached on the portfolio).
3. Match by name against property_name (watch for suffixed codes like "(tn97)" and typo'd cities). PATCH /api/assets/:id {espm_property_id}.
4. Expect true misses (Cortland Belmar). Record them in the completion report as "needs ESPM property creation or client confirmation" — do not block.

Stage 5 — Upload + index documents

Routing: per-asset docs → POST /api/assets/:id/files (multipart file + folder); portfolio-wide reference/helper files → POST /api/portfolios/files. Folder by type: Audits (energy/water audit reports), PCAs (PCAs, PCRs, surveys), Reference Data. Indexing is queued automatically for indexable MIMEs.

Hard-won mechanics (all hit on Cortland):

• 50MB API cap → split bigger PDFs with PyMuPDF (insert_pdf page ranges, "(Part N of M)" names). Lossless split beats lossy compression for source docs.
• curl -F breaks on commas in filenames → always quote: -F 'file=@"<path>"; type=<mime>' (returns empty-body failure otherwise).
• Portfolio files endpoint 500s on non-ASCII filenames (em-dash) → rename to ASCII first. It also 500s on text/markdown / application/json MIME in prod → send md/json/txt as text/plain (still indexes).
• Verify indexing: poll files.indexing_status until all indexed. On failed, check Railway deploy logs (project soapbox-platform, service soapbox-api, deploys from main) and retry with POST /api/assets/:id/files/:fileId/reindex. Known fixed bug: PDFs with NUL bytes in extracted text (fixed ac77a12).
• Count embeddings chunks at the end as a sanity check (embeddings table by asset/portfolio id).

Stage 6 — Verify + report

• SQL rollup: assets / audette_linked / espm_linked / files / indexed / chunks.
• Optional but recommended: Playwright pass on https://app.soapbox.build (service account login → workspace switcher → client org) to confirm the portfolio renders with Audette/ESPM badges; screenshot for the report.
• Completion report to the user: totals, per-asset link status, misses + follow-ups (ESPM gaps, unmatched docs, missing financial params), and what was NOT done (address mapping, LL/TT params if not collected).
• metadata.setup_complete stays false until the client verifies helper data; LL/TT allocation inputs (lease structure, metering config, jurisdiction) are collected in the review pass per the 2026-06-27 design spec when running the full conversational flow.

Idempotency

Before creating anything, check for an existing org/portfolio by name and existing assets by name + client tag. Resume from the first incomplete step; never re-upload a file whose name already exists under the same asset; never overwrite existing audette/espm links.