Legal Vault

The firm’s AI-ready document corpus. Indexed for hybrid retrieval (pgvector + tsvector + Reciprocal Rank Fusion) and surfaced to the Assistant, Workflows, Practice agents, Templates, and Bots.

Three ways to populate

Path	Where	When
Paste text	`Paste text` button on `/admin/legal/vault`	One-off briefs, memos
Upload file	`Upload file` button	PDF / DOCX / TXT / XLSX. Server extracts via pypdf, falls back to tesseract OCR for scanned PDFs, walks every sheet for xlsx
DMS import	`/admin/legal/dms`	Bulk — connect iManage / NetDocs, browse folders, import

Plus the email-to-Vault gateway for inbound correspondence and the Word add-in for ad-hoc text fragments.

Per-document fields

title             "Smith v. Jones — MTD (2024)"
classification    brief | precedent_contract | depo_transcript |
                  memo | exhibit | statute_notes | template |
                  correspondence | other
tags              ["delaware", "indemnity", "12(b)(6)"]
practice_area     m_and_a | litigation | tax | ip | family ...
source_matter_id  (optional — links to a matter)
inline_text       the body (auto-extracted for uploads)
acl_mode          "open" | "matter_team"

Ingest pipeline

Chunker splits into ~1200-char chunks with 200-char overlap (cap 2000 chunks per doc).
Embedder produces pgvector embeddings via the tenant’s embedding-task AI Provider.
Postgres tsvector is computed via generated column for keyword search.
ivfflat + GIN indexes for hybrid retrieval.

Status chip cycles pending_ingest → ingesting → ready (or failed with the reason). The Vault page auto-refreshes every 5 seconds while jobs are in flight.

Prerequisites

Must have at least one AI Provider with task_routing.embedding = true under AI providers. Without it, ingest goes to failed: no_embedding_provider. The Assistant and Workflows also need a task_routing.chat = true provider.

ACL modes

open (default) — anyone with the pollenix.legal entitlement and a legal role reads it.
matter_team — only members of the document’s source_matter_id team can read. See Matter-team ACLs.

Toggle at upload time via the Restrict to matter team checkbox.

Sources retrieved

The Vault is one of three knowledge sources the Assistant can consult. See Knowledge sources.