Lante

Problem

In many work settings, keeping multiple projects straight at once is hard — and it gets harder as each project grows in moving parts, stakeholders, and exceptions.

Most people do not live in a project-management tool. They live in email and WhatsApp, especially in LATAM. Updates, handoffs, and “where did we leave this?” moments scatter across threads instead of sitting in one place you can trust.

Lante ingests what people already send and receive on those channels, then maintains a current-state view per project — not just search, but “what is the status of this shipment / order / client right now?”

We focused the first domain on freight logistics on purpose: multiple orders in flight, protocol-heavy workflows, and sudden changes (delays, re-routes, customs holds) are normal daily work. That sector stress-tests whether the system can stay accurate when volume and volatility are both high.

Solution

Upload a WhatsApp export (or a PDF). Once it is ingested — and confirmed in Bandeja if the routing was ambiguous — you can ask, in Spanish, ¿Quién prometió la entrega del pedido #8890? and get citations pointing at specific messages, plus a timeline of the Friday promise, the slip, and the Monday re-promise. The seeded #8890 chat is one of the ambiguous ones: it spans several topics, so it does go through review.

The web PWA is the shipping client. A shared TypeScript package (apps/shared) is the API client for web (and the Expo app, which typechecks only). FastAPI serves the API. Docker Compose runs Postgres, Redis, Qdrant, FalkorDB, MinIO, the API, a Postgres-polling ingest worker, and a rules worker.

Walkthrough and private repo access are available on request — there is no public always-on URL. See How to see it.

Tech stack

Layer	What actually ships
Language / API	Python 3.11, FastAPI, Pydantic, SQLAlchemy, Alembic (12 revisions)
Web	Next.js 15, React 18, TypeScript; PWA via manifest.json, sw.js, install prompt
Shared client	apps/shared — typed fetch wrappers used by the PWA
Datastores	Postgres 16 (system of record), Redis 7 (rate limits and short-lived tokens — not the ingest queue), Qdrant (chunk vectors), FalkorDB (entity graph over the Redis protocol), MinIO (raw uploads)
LLM	Default gpt-4o-mini at extraction temperature 0; embeddings text-embedding-3-small (1536-d); optional per-workspace Qwen via DashScope
Eval / NLP extras	Hand-annotated JSONL gold sets, rapidfuzz, pdfplumber for PDFs
Mobile / extras	Expo 52 app typechecks in CI; Telegram bot module exists; telegram: false on every plan in packages/billing/tiers.py

Redis is in the stack. The worker does not consume a Redis queue — it polls the Postgres Job table.

Skills this demonstrates

Skill	Evidence in the repo
Applied NLP / extraction	Spanish LLM prompt + JSON schema, heuristic _mock_extract fallback, commitment regex overlay, rapidfuzz resolver (threshold 85)
Information retrieval	Hybrid retriever: Qdrant vector search + FalkorDB Cypher for commitment edges, lexical +0.1 fusion, citations with sender and timestamp
Evaluation design	93-row extraction gold/test split, 48 negatives, 20 phenomena (≥3 each), circular-fixture hygiene test, retrieval gold set of 21 questions with sender/date citation scoring
LLM product engineering	Tool-using ops copilot, deterministic synthesis for delivery-promise questions, bilingual input blocklist, Spanish refusals, optional NVIDIA NeMo topic/content guards (off by default)
Multi-tenant backend	require_workspace_access / WorkspaceDep; test_route_auth_coverage.py enumerates live routes; per-workspace payload check on every vector hit
Full-stack product	Inbox HITL, timelines, copilot UI with citation cards, audit CSV/PDF export, PWA
Production judgement	Rate limits fail open; embedding-dimension mismatch raises; demo mode refuses to boot without closed registration + read-only email; CI blanks LLM keys on the unit lane

This does not demonstrate statistical inference, forecasting, or training a model from scratch (that’s Sol).

Architecture

Two indexes on purpose. Vector search answers “what was said about freight.” Graph traversal answers “who promised what.” Retrieval queries both.

Layer	Choice	Why (and where)
Chunking	800 characters, 100 overlap	packages/ingestion/chunker.py
Extraction	gpt-4o-mini, temperature 0, extract_entities_es.md	Informal Colombian Spanish; heuristic path if the LLM is empty
Embeddings	text-embedding-3-small (1536-d); mock 384-d without a key	packages/indexing/embedder.py — collection dim is fixed at creation
Vector store	Qdrant	One shared collection. document_id uses a server-side filter; workspace_id is checked on each returned payload (see below)
Graph	FalkorDB	Cypher; smaller ops footprint than Neo4j
Entity resolution	rapidfuzz token_sort_ratio, threshold 85	“Fernando” / “Fernando Quintero” / “don Fernando”
Serving	FastAPI + Postgres-polling worker	No message broker; ingest is minutes-scale

Where tenant isolation actually happens on vector search. All workspaces share one Qdrant collection. document_id is pushed down as a server-side Filter, but workspace_id is enforced in application code — vector_writer.search() drops any returned point whose payload workspace_id does not match. Isolation holds, but two things follow honestly: another tenant's nearest neighbours can consume part of the top_k budget before being discarded, and the guarantee lives in one Python line rather than in the database. A payload index with a server-side must clause is the correct fix and is not done.

Human-in-the-loop is conditional, not universal. should_queue_for_review() sends a document to Bandeja when it splits into multiple segments, when more than one topic is detected, when a routing suggestion scores below 0.7 confidence, or when the workspace sets require_timeline_approval (default false). A single-topic, high-confidence upload routes straight to a timeline. The claim is that the ambiguous cases are queued — not that a human approves everything.

Measurement

Two harnesses are worth quoting. A third file, agent_eval.jsonl, has three cases and still mentions the old fixture’s Carlos / #4521 — it is not a published metric.

Extraction (offline)

Hand-annotated 93 rows (62 gold / 31 test) over 7 co_logistics_*.txt WhatsApp exports. Twenty phenomena, each tagged at least three times. 48 rows have empty entities and relations (deliberate negatives). Every row is "synthetic": true.

Required entities: 43 on gold, 63 across both splits. Relations: 8 total (4 per split).

Reproduced this session, no API key:

python scripts/run_extraction_eval.py --split gold --predictor heuristic

Predictor	Entity F1 (micro)	Relation F1 (micro)
Heuristic regex + mock extract (gold, 62 rows)	0.692 (P 0.771 / R 0.628; 27 TP, 8 FP, 16 FN)	0.143 (1 TP, 9 FP, 3 FN)
gpt-4o-mini, temperature 0 (gold, recorded in datasets/README.md)	0.933 (0.923–0.944 over 3 runs)	0.800 (same across those 3 gold runs)

The 9 false relations vs 1 true one is why the LLM path exists. On the same heuristic run, colombian_logistics_term entity F1 is 0.286 and negative_ack is 1.000; type confusion is Location→Person 3, Organization→Person 2.

LLM entity/relation numbers above are not re-run in this pass (they need OPENAI_API_KEY). They are the figures recorded with the corpus in packages/evaluation/datasets/README.md.

Retrieval (live stack)

retrieval_gold_set.jsonl — 21 questions covering all seven conversations. Scoring: fraction of expected_contains substrings in the answer, and whether a citation matches expected_citation_sender and the same calendar day as expected_citation_timestamp. Gate in scripts/run_eval.py: retrieval ≥ 0.80, citation ≥ 0.70. Needs Docker + a seeded workspace + an API key. No number from that gate is quoted here.

Honesty about the numbers

Relation F1 is under-powered. Eight relations total. On the test split, two LLM runs of the same prompt scored 0.571 and 1.000. Do not treat gold 0.800 as a stable capability number.

Single annotator, synthetic corpus. No inter-annotator agreement. Treat F1 as an upper bound. Real anonymized exports would set "synthetic": false in the same schema.

The old retrieval gold set was circular. packages/evaluation/gold_set.jsonl still exists to reproduce old numbers; it is not the default gate.

Failure modes

Observed on this corpus / this code:

1. Hedges. “si acaso llega el viernes, todavia no es seguro por el clima” is tagged hedged_noncommitment (3 rows). commitment_extractor.py still fires on entrega-adjacent regex. Inventing a promise is worse than missing one.
2. Superseded promises are appended. #8890 Friday then Monday: both edges can exist. valid_from is set on relations (temporal_edges.py); nothing marks the first as void.
3. Accents. Gold names are verbatim (trancón, Julián). Matching normalizes accents in eval; a “fix” to gold strings would still be wrong.
4. Heuristic type confusion. Capitalized Spanish tokens default to Person in _mock_extract — hence Location→Person / Organization→Person on the baseline.
5. Implicit subjects. “hay un trancon verraco en la via al Llano” is annotated with a location and no relation, because the extractor sees msg.body, not sender: body.
6. The deterministic answer is tuned to this corpus. answer_synthesis.py fires on any delivery-promise question naming an order of three or more digits, but the two-date sentence (“first Friday, then rescheduled to Monday”) is keyed on the literal words viernes and lunes. It is a correct, cheap answer for the seeded chats and would need real date parsing to generalise.

Production concerns

Concern	Implementation
Multi-tenant isolation	Path/body workspace_id routes must call the membership helper. tests/unit/test_route_auth_coverage.py walks the live app. Deny-behavior is in test_tenant_isolation.py (requires_infra, not the offline 431)
Rate limiting	Redis fixed-window: login 5/min/IP, register 3/hour/IP, LLM/copilot 30/min per identity (/v1/consulting/chat included). Fails open if Redis is down
Prompt injection	Two layers on the copilot path: a 500-character cap plus a bilingual blocklist (security_blocklist.py — Spanish and English instruction-override, system-prompt extraction, code fences, SQL-injection asks), then an optional NeMo NIM topic check (nemo_guardrails_enabled defaults false) and an output-safety pass. A separate four-phrase English check sits in guardrails.py. Refusals are Spanish
Cost	Per-workspace OpenAI vs Qwen. Copilot 30/min. Delivery-promise questions can skip the LLM via answer_synthesis.py
Migrations	Alembic, 12 revisions; migrate container must exit 0 before api starts
Embedding dim	Mock 384-d vs OpenAI 1536-d; embedding_dimension() must match the Qdrant collection
Demo safety	ENVIRONMENT=demo will not boot unless registration is closed and DEMO_READONLY_EMAIL is set. That user cannot mutate; copilot and /v1/audit/export are allowlisted

431 passing tests in pytest tests/unit with LLM keys blanked (CI job test). One more collected test is skipped; 37 requires_infra tests are a separate job.

Honest scope

Not in this project: model training, statistical inference, forecasting, a live Creem checkout, an observability stack, a mobile store build, or a live email/WhatsApp inbound relay (webhook secrets exist so unsigned requests can be rejected).

How to see it

1. This page — use Request a walkthrough below to schedule a live session. Screenshots/recording after local capture (docs/SCREENSHOTS.md in the repo).
2. On request — seeded workspace, demo script: login → feed → inbox → copilot on #8890 → timeline + export. Request a walkthrough.
3. Private repo — on request. Offline reproduce:

pip install -e ".[dev]"
python scripts/run_extraction_eval.py --split gold --predictor heuristic
python -m pytest tests/unit -q

Full stack: docker compose --profile app up -d --build, then scripts/seed_public_demo.py.