Features

A workspace (tenant) holds members and RagSpaces; each RagSpace owns its documents, chunks and LLM/embedding configuration.

Soft deletion of a space (deleted_at); hard purge on DELETE.

rag_space is duplicated on every Chunk and filtered by a double SQL clause; access re-checked before Core calls; anti-enumeration (404).

Owner-level scoping inside a workspace is optional (see "owner scopes").

Accounts created by one-shot invitation links (SHA-256 hashed) or admin bootstrap; anti-enumeration password reset; no public sign-up.

Sending e-mail requires an SMTP configuration; by default messages go to the console.

Mandatory TOTP second factor (django-otp) for every account, no opt-out flag; django-axes lockout, anti-replay, backup tokens, OTP-protected Django admin.

TOTP and static backup tokens only; no WebAuthn or SMS.

Roles owner/admin/member/viewer; viewer read-only; owner/admin manage members, RagSpaces, LLM settings and audit; superuser is orthogonal (shared corpus, config).

The last owner cannot be removed; an admin cannot remove an owner.

Cross-workspace governance (/api/v1/platform/*, /platform-admin screen): users, workspaces, owners, diagnostics; no content access; Django admin behind OTP.

The superuser has no implicit access to content.

Screens /workspace-admin (members, RagSpaces, invitations), /admin-core (Core config, document capabilities), /ragatm-admin (shared corpus).

Gateway probes from the console are off by default.

Shared document layer across workspaces (domain_corpus_* tables), superuser-administered, synchronous ingestion, bi-source retrieval with isolation preserved.

The shared corpus does not take part in answers by default (calibrated admission).

Upload via /api/v1/documents/upload/ with content sniffing; default formats PDF, DOCX, PPTX, XLSX, CSV, TXT, MD, HTML, JPEG/PNG/TIFF/WEBP images; 50 MB limit; sha256 dedup.

Images are accepted at upload but fail extraction in Compact (no OCR). Upload limit: 50 MB in Compact (25 MB default elsewhere in the documentation).

Ingestion through a Celery/Redis worker; Compact sets RAG_INGEST_ASYNC_DEFAULT=1; OCR/Excel/PPTX/DXF/IFC force async; broker down → 503 dependency_unavailable.

Single worker, concurrency 2 in Compact.

Dedup by (rag_space, sha256) excluding FAILED; a duplicate creates no file or task; the same file can be imported in another space.

No particular limit noted.

Hard immediate document deletion (409 while ingesting); RagSpace DELETE = irreversible purge (documents, chunks, vectors, threads) keeping ActionAuditEvent.

No recycle bin.

Per-RagSpace reindex jobs on the reindex queue, retry of failed documents, strategies auto / force_ocr / extended_timeout; no implicit reindex on model change.

Changing the embedding model requires an explicit reindex.

Document: filename, sniffed content_type, size, sha256, document_date, status, JSON meta, embedding provider/model, reindex_required…; Chunk: chunk_index, page_number, pipeline_id/family, meta (per-format locators).

The document "confidentiality" field is inert (read by no code).

pypdf (+pdfminer/pdfplumber fallbacks), DOCX via stdlib, python-pptx, openpyxl, csv/html stdlib; scanned PDF without OCR → explicit failure pdf_scanned_requires_ocr.

DOCX: headers, footers, footnotes, comments and text boxes not extracted; PyMuPDF engine absent (AGPL).

Deterministic per-page chunking: 2500 chars / 250 overlap; governed spaces use min(2500, model cap) → 1200/180 with Granite; dedicated Excel (40 rows) and PPTX (one slide = one unit) strategies.

The embedding model’s 1,200-character limit bounds chunk size.

Single provider ai_gateway; logical model fr-en-retrieval (granite-embedding:278m, 768 dims); identity (pipeline_id, digest, identity_verified) enforced; checks E101/E102; refuses to index/query with a different pipeline.

granite profile qualified "experimental"; the legacy 768 profile is not installed in Compact.

pgvector cosine leg filtered by rag_space and pipeline_family, subject to the RAG_MIN_SCORE threshold.

Similarity threshold inherited from another model: it may empty the vector leg with granite (see retrieval threshold).

PostgreSQL FTS to_tsvector/ts_rank, config simple by default, query plan from most to least selective; stored tsvector column behind a flag (off).

The stored-vector path requires a migration and RAG_FTS_CONFIG=french.

merge_hybrid fuses FTS and vector legs by Reciprocal Rank Fusion (K=60) with conditional lexical boost on strong identifiers; RAG_HYBRID_ENABLED=1 by default.

No particular limit noted.

Candidate pool ×3, max 2 per group (document) before decision; on by default.

No particular limit noted.

Passage reranking via POST /v1/rerank (TEI bge-reranker-v2-m3 behind the Gateway); OFF in code defaults, ON in Compact (pool 20, timeout 20 s); on failure, counted fallback to the canonical RRF order (fail_open_rrf).

Fallback = hybrid RRF order (not a lexical rerank); CPU latency ≈ 1.5 s per pair; two reranking settings mutually exclusive.

should_answer: min_sources, lexical threshold, vector threshold RAG_MIN_SCORE (code default 0.76, calibrated for nomic); hard refusal only when zero chunks, else partial/low_confidence mode.

Documentation drift on the default value (0.20 vs 0.76); 0.76 not calibrated for granite.

All generation goes through the Gateway's POST /v1/generate (logical model compact-generation-8 = qwen2.5:1.5b-instruct in Compact); per-RagSpace model; direct Ollama path reserved for tests.

The 1.5 B model demonstrates the path, not answer quality; unregistered or uninstalled models refused.

/chat/stream/ emits SSE (meta → sources → token(s) → done, + error); with AI_GATEWAY_STREAMING_ENABLED=1 (Compact) fragments come from /v1/generate/stream (NDJSON); at 0, single-block pseudo-stream.

Some documentation pages still say "no real streaming": documentation drift, the Compact configuration enables native streaming.

Sources v1 schema (rank, document_id, chunk_id, score, filename, page_number, snippet ≤200); answer contract separating retrieved / prompted / actually cited sources; 'Sources:' line only for real citations.

Citation correctness depends on the model honouring the [n] markers.

Citation guards at prompt head; no_answer without any LLM call when no source; 'off-documents' Mode B disabled; FR/EN abstention detection; no fabricated answer on failure (typed errors).

Cross-encoder answerability measured but not promoted; anti-over-abstention retry evaluated, not implemented.

Persisted ChatThread/ChatMessage, private to their author; hierarchical memory (12-message window, summary beyond 18) injected only into the prompt; bounded thread search (3 s).

The memory summary uses the global model, not the space model (known limit).

Short synchronous / long Celery map-reduce summaries persisted as ChatMessage; shared-corpus document summaries on by default (not qualified).

Corpus summaries not qualified.

Asynchronous CorpusAnalysisJob (analytics queue) with lifecycle, progress, cancellation; RagSpace access re-checked on every operation.

No particular limit noted.

Static product help, Excel analytic guard (deterministic refusal), document listing: short-circuit RAG without any LLM call.

No particular limit noted.

Preview routed by preview_kind: pdf (iframe #page=N), spreadsheet, document (DOCX), presentation (PPTX), text, unsupported; always on.

CSV deliberately unsupported by the viewer.

Thumbs-down → diagnosable incident (lot 1 on); incident export/import and pseudonymisation behind flags (off).

No particular limit noted.

RagQueryLog (90 d), ActionAuditEvent (365 d), LlmEgressAuditEvent / EgressPolicyChangeLog (365 d with a check-enforced floor), LlmSettingsChangeLog; periodic purge.

The RAG query log stores question and answer in clear text (conversational data).

manage.py check (llm_router.*, config.*, rag_core.* checks), diagnose.sh (non-inferential), retrieval-diagnostic.sh, contract_check.py, resilience_check.py, offline-check.sh, e2e_qualification.py.

Inferential gateway probes are never called automatically.

backup.sh (custom pg_dump, media archive, configs, model manifest, SHA256SUMS) and restore.sh into an isolated instance; secrets, Redis and model weights excluded.

Documents being ingested at backup time must be reindexed after restore; secrets must be backed up separately.

next-intl without route prefix; locales fr (default) and en; ragatm_locale cookie + account preference; 21 symmetric namespaces under test.

Roll-out to application screens in progress per the documentation.

~128 DRF routes under /api/v1/ (auth, bootstrap, capabilities, workspaces, invitations, rag-spaces, documents, chat, analyses, prompts, audit, core, admin, platform, tabular); contract shared/contracts/openapi.yaml 1.0.0; cookie token + CSRF or Authorization: Token.

Hand-maintained OpenAPI contract, not automatically checked against the code.

DRF User/Anon throttles in production: user 120/min, anon 30/min, auth 20/min, mfa 10/min, feedback 10/min, thread search 60/min, invitations 20/10 per min.

Anonymous bucket shared by all clients behind the Next proxy (known limit).

prod.py: 1-year HSTS + subdomains + preload, nosniff, X-Frame-Options SAMEORIGIN, Referrer-Policy same-origin, Secure/HttpOnly/SameSite=Strict cookies; TLS expected at a reverse proxy.

No TLS inside the Compact stack; no Content-Security-Policy in the settings reviewed.

Middleware relays a conformant X-Request-ID ([A-Za-z0-9._:-]{1,128}) or generates a UUID4; propagated to the Gateway (generation, embeddings, rerank) and SSE streams; present on all responses.

No particular limit noted.

503 ai_gateway_unavailable (+Retry-After, retryable); 502 ai_gateway_error (not retryable); 503 dependency_unavailable (redis|broker); same payload in the SSE error event; never a fabricated answer.

No particular limit noted.

max_document_count and max_document_bytes copied from platform defaults; usage aggregated on demand; enforcement on by default.

No unlimited mode.

Read-only superuser screen exposing each capability's real state (ACTIVE/DISABLED/UNAVAILABLE/UNKNOWN); Gateway /health/* probes closed by default.

Service states shown "UNKNOWN" until probes are enabled.

Structured trace per query_rag (observable degradations), bounded ops counters (rag_reranker_*), /ops/metrics outside /api/ with optional token, non-inferential /api/health/.

No particular limit noted.

Every LLM call needs an LlmExecutionContext with a resolved policy; any non-local provider is refused; allow_external_fallback is traced but no external provider exists.

Granularity stops at the document space; inert confidentiality field.

Partitioning several owners inside one workspace plus per-RagSpace ACL; implemented end-to-end but off by default.

Not enabled in Compact; scope transfer from the platform console is preview-only.

answer_language preference (auto/fr/en) driving generated-answer language, distinct from UI locale; flag off in code.

Off by default in the Open code; Compact does not set it.

Chat action routing (apps.actions); mutations never triggered by phrasing; off by default.

No particular limit noted.

DuckDB engine computing on the source xlsx/csv file (never the LLM, never chunks); closed schema, row/column/time bounds; off by default and in Compact.

LLM planner and narration not wired.

Database-stored configuration overrides, inert by default; editing and runtime effect behind strict flags (off).

No particular limit noted.

.eml ingestion with attachments behind a flag and per-space scope (off); Outlook .msg unsupported in the Open edition (GPL parser dropped, code msg_unsupported_in_open_edition).

Attachments: no vision or structured extraction; IFC/DWG unsupported.

Query rewrite/keywordisation in explanatory mode; LLM rewrite disabled after A/B measurement (no gain, doubled P95).

No particular limit noted.

Legacy rerank loading a model in-process (rerank/torch extra); off; never locked for the Compact image.

Exclusive with reranking through the gateway.

No mailbox connector in the Open edition (explicit code comment).

No particular limit noted.

The three Gateway URLs (generation, embeddings, rerank) point to http://ai-gateway:8100; empty AI_GATEWAY_POOL_HOSTS ⇒ direct route; nominal Compact profile.

No particular limit noted.

Compact: db (pgvector), redis, backend (gunicorn 2×4), worker (Celery), beat, frontend (Next), ai-gateway, ollama, reranker (TEI) on 5 networks, 4 internal; per-service memory budgets.

Single host; no multi-machine deployment verified.

Public queues celery, reindex, maintenance, analytics consumed by one worker (concurrency 2, max-tasks-per-child 50); dedicated ingest_visual/audio/ifc/dwg queues defined but without consumer in Compact.

Sequential ingestion at workstation scale.

Gateway 2 concurrent generations (503 beyond), Ollama NUM_PARALLEL=1 / 2 loaded models, worker concurrency 2, gunicorn 2 workers × 4 threads, 600 s timeout; no Gateway-side queue.

Load beyond one concurrent generation not verified.

Originals are written to a local volume (media): storage tied to a single host.

No object or shared storage.

Standalone HTTP router relaying the 19 gateway-api/1.0 routes to 1..N Gateways (SINGLE_NODE, WEIGHTED_ROUND_ROBIN, LEAST_INFLIGHT), node registry with allowlist, model governance (no substitution, digest identity); absent from Compact; all-or-nothing switch (llm_router.E008).

ASR/vision not qualified through the Pool; full RAG through the Pool not demonstrated in the repositories.

Per-node max_inflight → AT_CAPACITY (refusal, no queue); CLOSED/OPEN/HALF_OPEN breaker on transport failures; soft suspension on 503+Retry-After; failover only before any byte is sent (connect errors); process-local counters.

Process-local state, no distributed coordination; no queueing, priorities or rate limiting.

Job→node affinity store (none/memory/sqlite/redis) for asynchronous ASR jobs; the Pool's only Redis use; backend none in Compact (audio routes refused).

No particular limit noted.

README: load beyond one concurrent generation unverified; Pool: high-concurrency throughput not re-measured; sequential ingestion.

No particular limit noted.

Multi-node configuration prepared (EXAMPLE_ONLY); MULTINODE_HARDWARE_QUALIFIED = NO; LIVE_MULTI_GPU NOT_EXECUTED.

No particular limit noted.

No autoscaling or multi-replica coordination feature: single-process Pool, process-local state; only ASR affinity is shared.

No particular limit noted.