Riverstone

How the analysis reads your documents and checks them against this checklist: the document types, extraction rules, and how findings are tagged to checklist items.

# Memo Agent — Data Room Ingestion Schema
# Version: 0.1 (draft)
# Used by: Memo Drafting Agent (Stage 1 — ingestion)
#
# Defines: how each document in a Google Drive data room gets parsed
# into structured claims, what document types we recognize, what we
# expect to find (so we can flag gaps), and how financial models get
# special treatment so company projections don't leak in as facts.

meta:
  version: "0.1"
  last_updated: "2026-05-06"
  owner: vendor   # Technical schema; partners review but don't typically edit.

# ---------------------------------------------------------------------------
# Provenance — every claim from the data room carries this by default
# ---------------------------------------------------------------------------

provenance_default:
  source_type: company_stated
  rationale: |
    Every claim extracted from the data room is treated as
    "company-stated" — the company has asserted it; it has NOT been
    independently verified. The agent must never elevate a
    company-stated claim to a fact in the memo. Verification belongs
    to Stage 2 (external research) and Stage 3 (partner Q&A). This
    is the single most important guardrail in this schema.

# ---------------------------------------------------------------------------
# Document types
# ---------------------------------------------------------------------------

document_types:
  - id: pitch_deck
    name: "Pitch deck"
    typical_formats: [pdf, pptx, gslides]
    typical_content: [problem, solution, market, traction, team, ask]
    extraction_priority: high

  - id: financial_model
    name: "Financial model / projections"
    typical_formats: [xlsx, gsheets]
    typical_content: [historical_metrics, projections, assumptions, unit_economics]
    extraction_priority: high
    special_handling: financial_extraction   # see below

  - id: cap_table
    name: "Cap table"
    typical_formats: [xlsx, gsheets, pdf]
    typical_content: [ownership, option_pool, prior_rounds, preferences]
    extraction_priority: high

  - id: founder_bios
    name: "Founder bios / resumes"
    typical_formats: [pdf, docx]
    typical_content: [education, prior_roles, exits, publications]
    extraction_priority: high

  - id: customer_reference
    name: "Customer reference / case study"
    typical_formats: [pdf, docx, gdocs, video, audio]
    typical_content: [customer_name, use_case, outcomes, willingness_to_reference]
    extraction_priority: high

  - id: product_demo
    name: "Product demo"
    typical_formats: [video, mp4, mov]
    typical_content: [product_walkthrough, ux, capabilities]
    extraction_priority: medium
    note: "Transcribe audio; extract feature list and key claims."

  - id: technical_doc
    name: "Technical architecture / spec"
    typical_formats: [pdf, docx, gdocs, md]
    typical_content: [architecture, stack, ip, moat_claims]
    extraction_priority: medium

  - id: market_analysis_internal
    name: "Company-prepared market analysis"
    typical_formats: [pdf, docx, gdocs, pptx]
    typical_content: [tam_sam_som, competitors, trends]
    extraction_priority: medium
    note: |
      Treat with extra skepticism. Company-prepared analyses tend to be
      selectively sourced. Flag every claim for verification in Stage 2.

  - id: prior_round_terms
    name: "Prior round terms / SAFE / agreements"
    typical_formats: [pdf, docx]
    typical_content: [valuation, investors, rights, preferences]
    extraction_priority: high

  - id: hiring_plan
    name: "Hiring plan"
    typical_formats: [xlsx, gsheets, pdf]
    typical_content: [roles, timing, comp_bands]
    extraction_priority: medium

  - id: roadmap
    name: "Product roadmap"
    typical_formats: [pdf, pptx, gsheets]
    typical_content: [milestones, timing, features]
    extraction_priority: medium

  - id: whiteboard_or_sketch
    name: "Whiteboard photo / strategy sketch"
    typical_formats: [jpg, png, heic]
    typical_content: [informal_diagrams, strategy_notes]
    extraction_priority: low
    note: "OCR plus visual interpretation. Often partial; flag low confidence."

  - id: press_coverage
    name: "Press coverage / clips"
    typical_formats: [pdf, link, gdocs]
    typical_content: [headlines, publication, date, tone]
    extraction_priority: low
    note: "Useful as signal, not as fact."

  - id: other
    name: "Other / unclassified"
    typical_formats: [any]
    extraction_priority: low
    note: "Catch-all. Surface for partner review rather than guessing."

# ---------------------------------------------------------------------------
# Expected documents — drives gap analysis
# ---------------------------------------------------------------------------

expected_documents:
  description: |
    What we'd typically expect to find in a complete data room. Items
    not present become entries in the gap_analysis output. Absence
    isn't necessarily a flag — but it should always be visible.
  expected:
    - { id: pitch_deck,         criticality: required }
    - { id: financial_model,    criticality: required }
    - { id: cap_table,          criticality: required }
    - { id: founder_bios,       criticality: required }
    - { id: customer_reference, criticality: expected }
    - { id: product_demo,       criticality: expected }
    - { id: prior_round_terms,  criticality: expected }
    - { id: hiring_plan,        criticality: optional }
    - { id: roadmap,            criticality: optional }

# ---------------------------------------------------------------------------
# Per-document extraction format
# ---------------------------------------------------------------------------

document_record:
  fields:
    - { name: id,                  type: string,   required: true,  note: "Stable per-data-room ID." }
    - { name: file_name,           type: string,   required: true }
    - { name: file_format,         type: string,   required: true,  note: "pdf, xlsx, gslides, mp4, etc." }
    - { name: drive_path,          type: string,   required: true }
    - { name: drive_file_id,       type: string,   required: true,  note: "Google Drive file ID for traceability." }
    - { name: detected_type,       type: string,   required: true,  note: "ID from document_types." }
    - { name: type_confidence,     type: enum,     required: true,  values: [low, medium, high] }
    - { name: document_date,       type: date,     required: false, note: "Author date if found; else file modified date." }
    - { name: page_or_slide_count, type: integer,  required: false }
    - { name: parse_status,        type: enum,     required: true,  values: [parsed, partial, failed, skipped] }
    - { name: parse_notes,         type: string,   required: false, note: "Why partial/failed (scanned, password-protected, etc.)." }
    - { name: summary,             type: string,   required: true,  note: "1–3 sentence summary of the document's purpose and content." }
    - { name: claims,              type: list,     required: false, note: "List of claim_record entries — see below." }
    - { name: financial_extract,   type: object,   required: false, note: "Only for financial_model; see financial_extraction." }

# ---------------------------------------------------------------------------
# Claim record — every meaningful assertion extracted from the data room
# ---------------------------------------------------------------------------

claim_record:
  description: |
    A discrete assertion extracted from a document. Each claim is
    tagged with provenance, the section of the memo it likely feeds,
    and a verification status the research stage will update.
  fields:
    - { name: id,                  type: string,  required: true,  note: "Stable claim ID, e.g. 'pitch_deck_001_claim_03'." }
    - { name: text,                type: string,  required: true,  note: "The claim, phrased neutrally — strip marketing voice." }
    - { name: claim_type,          type: enum,    required: true,  values: [metric, market_size, competitive_position, customer_evidence, team_credential, product_capability, financial_projection, other] }
    - { name: source_document_id,  type: string,  required: true }
    - { name: source_location,     type: string,  required: false, note: "Page, slide, cell — wherever the claim lives." }
    - { name: source_type,         type: enum,    required: true,  values: [company_stated, third_party_cited_by_company, customer_quoted_by_company] }
    - { name: feeds_memo_section,  type: list,    required: true,  note: "Memo sections this claim is likely relevant to." }
    - { name: verification_status, type: enum,    required: true,  values: [unverified, verification_attempted, verified, contradicted, unverifiable] }
    - { name: verification_notes,  type: string,  required: false, note: "How and where it was verified, or why not." }

# ---------------------------------------------------------------------------
# Financial model — special handling
# ---------------------------------------------------------------------------

financial_extraction:
  rationale: |
    Financial models are dense with numbers that look like facts. They
    are not. Every projection rests on assumptions, and the assumptions
    are where the actual judgment lives. The agent extracts both, links
    each metric to its driving assumptions, and surfaces aggressiveness
    versus historical actuals where comparison is possible. The agent
    does NOT critique the model — it makes the assumption layer visible
    so the partner can.

  metrics_to_extract:
    - { id: arr,             name: "ARR",                   units: currency, historical: required, projected: required }
    - { id: arr_growth,      name: "ARR growth rate",       units: percent,  historical: required, projected: required }
    - { id: gross_margin,    name: "Gross margin",          units: percent,  historical: required, projected: required }
    - { id: net_margin,      name: "Net margin",            units: percent,  historical: optional, projected: optional }
    - { id: burn,            name: "Net burn (monthly)",    units: currency, historical: required, projected: required }
    - { id: runway,          name: "Runway (months)",       units: months,   historical: required, projected: required }
    - { id: cac,             name: "CAC",                   units: currency, historical: optional, projected: optional }
    - { id: ltv,             name: "LTV",                   units: currency, historical: optional, projected: optional }
    - { id: ltv_cac,         name: "LTV:CAC ratio",         units: ratio,    historical: optional, projected: optional }
    - { id: payback_months,  name: "CAC payback (months)",  units: months,   historical: optional, projected: optional }
    - { id: nrr,             name: "Net revenue retention", units: percent,  historical: optional, projected: optional }
    - { id: gross_retention, name: "Gross retention",       units: percent,  historical: optional, projected: optional }
    - { id: headcount,       name: "Headcount",             units: count,    historical: required, projected: required }

  metric_record:
    fields:
      - { name: metric_id,           type: string }
      - { name: value,               type: number_or_range }
      - { name: period,              type: string,   note: "e.g., 'FY2025', 'Q3 2025', 'projected EOY 2027'." }
      - { name: is_projection,       type: boolean }
      - { name: cell_reference,      type: string,   note: "Sheet name + cell, for traceability." }
      - { name: driving_assumptions, type: list,     note: "List of assumption_record IDs this metric depends on." }

  assumption_record:
    fields:
      - { name: id,                     type: string }
      - { name: text,                   type: string,  note: "The assumption, stated plainly. e.g., 'Sales rep ramps to full quota in 4 months.'" }
      - { name: cell_reference,         type: string }
      - { name: assumption_type,        type: enum,    values: [growth_rate, conversion_rate, churn_rate, headcount_plan, pricing, cogs, opex_ratio, other] }
      - { name: aggressiveness,         type: enum,    values: [conservative, in_line_with_history, optimistic, aggressive, unmarked], note: "Compared to historical actuals where available; 'unmarked' when no historical comparison possible." }
      - { name: partner_attention_flag, type: boolean, note: "True when the assumption is materially more aggressive than historical actuals." }

  rules:
    - "Never quote a projected number without surfacing its driving assumptions in the memo or appendix."
    - "Mark every projection as 'projected' in the memo; do not blur the historical/projected line."
    - "Flag assumptions materially more aggressive than historical actuals; do not editorialize on whether they are achievable."
    - "The agent does not score the model. It surfaces structure for partner judgment."

# ---------------------------------------------------------------------------
# Gap analysis — what's expected but missing
# ---------------------------------------------------------------------------

gap_analysis:
  description: |
    Run after all documents are ingested. Compare detected types
    against `expected_documents`. Output the list of expected items
    that were not found, with criticality and implication.
  output_record:
    fields:
      - { name: missing_type,            type: string }
      - { name: criticality,             type: enum,   values: [required, expected, optional] }
      - { name: implication,             type: string, note: "What this absence means for the memo (e.g., 'Without a financial model, projections section will be partner-driven only.')." }
      - { name: partner_recommendation,  type: string, note: "What the agent recommends asking the company for." }

# ---------------------------------------------------------------------------
# Ingestion output — the artifact handed to Stage 2 (research) and Stage 4 (drafting)
# ---------------------------------------------------------------------------

ingestion_output:
  fields:
    - { name: data_room_id,         type: string,    required: true }
    - { name: ingestion_date,       type: datetime,  required: true }
    - { name: documents,            type: list,      required: true,  note: "List of document_record entries." }
    - { name: claims,               type: list,      required: true,  note: "Flattened list of all claim_record entries across documents." }
    - { name: financial_summary,    type: object,    required: false, note: "Aggregated financial_extraction across the model(s)." }
    - { name: gaps,                 type: list,      required: true,  note: "List of gap_analysis entries." }
    - { name: partner_review_notes, type: string,    required: false, note: "Free text for the agent's overall observations — what surprised it, what's missing, what looked unusual." }

The schema is the active configuration from Settings → Memo agent. Editing the schema itself is admin-only.