Lekha documentation

Reuses the one brand voice your workspace already set for Socie — tone, audience and an avoid list — rather than keeping a second copy that could drift.

whatsapp_message, social_caption, product_description, email, sms, ad_headline and website_section, each with a real word or character limit that is now measured, not just described in the prompt. Draws source material from your recorded facts by default. Needs an LLM key; the fact store does not.

Needs: LLM_API_KEY. The knowledge base and the length check work without one.

Hinglish, English, Hindi, Marathi, Gujarati, Tamil, Telugu, Kannada, Bengali, Punjabi and Malayalam — named individually, because a model told only 'Hindi' produces transliteration nobody actually speaks.

Needs: LLM_API_KEY.

Holds numbers, prices and proper nouns fixed while making the rest read as a native speaker would write it. The avoid list is deliberately not re-checked across a language change — a Hindi draft can't be honestly checked against terms you wrote in English.

Needs: LLM_API_KEY.

The guard re-checks that the rewrite introduced no fact the original lacked, and a word-level diff reports any price, number or proper noun the rewrite dropped — the guard catches additions and this catches deletions.

Needs: LLM_API_KEY.

Shorten, expand, simplify, fix grammar, more formal, more casual — each returns what it dropped. There is deliberately no action meaning 'make this pass review': the guard runs again over every rewrite's own output.

Needs: LLM_API_KEY.

A workspace's own guideline rules, beyond the brand voice's avoid list: forbidden-term and required-phrase rules are pattern-matched — no model, no credential, the same check the avoid list already uses — and persisted per-guideline so 'which rule keeps failing' is a query rather than a re-read of blobs. tone_style rules cannot be pattern-matched by definition ('always write in second person' is not a regex) and are skipped and named as skipped, never silently passed, when no model is configured.

Needs: LLM_API_KEY for the tone/style half only — forbidden-term and required-phrase rules work without one.

URL extraction (through the same SSRF-guarded fetch Khoji's audits use) and pasted-text sources both work, one guarded generation call per target format, all conditioned on the same extracted source. Each output carries a source_excerpt — the source sentence it overlaps with most by word-overlap — so a derived caption is traceable back to a specific passage. Audio and video sources are refused by name: paste a transcript as source_type='text' instead.

Needs: LLM_API_KEY. Extracting the source itself needs no credential.

Platform-aware captions with hashtags, guarded and length-checked the same way every other draft is. image_ref is passed to the model as text context only — there is no vision call, so a caption is not actually grounded in what the photo shows; that gap is real and unbuilt.

Needs: LLM_API_KEY.

Single-item and row-by-row bulk generation, each guarded against inventing a material, ingredient or spec the row did not supply. Bulk rows arrive as a JSON array in the request body, not a CSV upload — a deliberate simplification named honestly. Capped at 50 rows per batch, with partial-failure handling: one bad row is marked failed and the rest of the batch still completes.

Needs: LLM_API_KEY.

Subject, body and CTA in one call; subject-only regeneration is a separate short call so A/B variants can be refreshed without redrafting the body. The guard also checks for unverified promotional superlatives ('guaranteed', 'the best', 'world-class') on top of the usual checks.

Needs: LLM_API_KEY.

Outline-first, then one model call per section, then a meta description and FAQ — chained so a 1200-word article costs roughly what its length costs. The draft runs through the same claim guard as every other Lekha output, plus a new pattern that catches uncited 'studies show' statistics.

Needs: LLM_API_KEY. Capturing the brief and hand-editing a draft afterward need no credential.

Reuses Khoji's own on-page checklist and its keyword-placement scorer rather than rebuilding either. A URL is fetched through the same SSRF-guarded client Khoji's audits use; pasted body copy is scored directly with no fetch at all. Every score carries checked_at so it never implies a live rank. No credential, no vendor.

A flag per finding, a decision that carries who decided and when, and a document that cannot reach approved while a flag on its current version is open. Needs no model — the same offline pattern-match guard Patri already enforces a hold with.

Every version is kept, never overwritten. A word-level diff names what a later version dropped, and a restore writes a new version rather than rewinding — the history between never disappears. No model call: diffing plain text doesn't need one.

The three states work, with approval gated on flag decisions and a queue of what's waiting. Inline comments (anchored to a position, resolvable) and per-role permissions — viewer, editor, approver, admin — now sit beside the workflow. The gate is opt-in per workspace: no roles configured means no restriction, the right default for a one-person workspace, and it turns on the moment a workspace assigns anyone a role. A comment or status change notifies everyone else with a role via the existing in-app bell — no credential.

Persona CRUD, with a 'reviewed' gate before a persona is treated as ready to use broadly, works with no credential; deleting one retires rather than deletes so a document generated against it keeps a resolvable reference. /write accepts a persona_id and blends its pain points and preferred tone into the same generation pipeline and the same guard — not a second code path.

Needs: LLM_API_KEY, the same requirement multi-format-generator already carries. Persona CRUD needs none.

A first-party copy-strength checklist (CTA presence, hook strength, specificity, length-for-format, filler-word count), genuinely deterministic and never called a prediction — every response says outright that there is no real outcome data behind it, unlike a vendor score dressed up as machine learning. Scoring and ranking several drafts against each other work with no credential.

Needs: LLM_API_KEY, only for the boost rewrite targeted at the lowest-scoring factors.

One key message derived from the brief, then one generation call per asset type conditioned on that same message rather than each asset inventing its own summary. A real consistency check compares every asset's own offer numbers against the brief's declared key facts, or against each other when none were declared, and flags the ones that disagree.

Needs: LLM_API_KEY for generating assets. Creating a campaign works without one (the key message falls back to a plain concatenation).

Cannot be built in-house — checking against 'everything published on the web' needs a web-scale index nobody here has. The full local half is real: request storage, response parsing defensive against an undocumented provider shape, and a fix for a false-positive problem — marking a match 'acceptable boilerplate' carries that acknowledgement forward automatically onto the same URL on a later check.

Needs: ORIGINALITY_API_KEY (Originality.ai). Unset on this deployment, so every check refuses with a 503 naming it.

Uses the same provider account plagiarism-originality-check uses (Originality.ai's scan/ai endpoint), since telling AI-generated text from human text needs a classifier trained for exactly that, not a language model guessing about its own kind. No accuracy percentage is ever stated in a response. A provider failure is recorded on the row rather than surfaced as a 500.

Needs: ORIGINALITY_API_KEY (Originality.ai). Unset on this deployment, so every check refuses with a 503 naming it.

Record a fact the guard may treat as evidence. A testimonial needs an attribution.

{
  "category": "pricing",
  "label": "2BHK deposit",
  "body": "Security deposit is ₹50,000, refundable."
}

{
  "id": "f-102",
  "category": "pricing",
  "label": "2BHK deposit",
  "body": "Security deposit is ₹50,000, refundable.",
  "is_active": true
}

Draft copy from a brief, checked before it's handed back.

{
  "brief": "New 2BHK listing in Kandivali, ready to move in",
  "format": "social_caption",
  "language": "hinglish",
  "variations": 1
}

{
  "copy": "2BHK ready to move in — Kandivali mein aapka naya ghar. Deposit sirf ₹50,000. DM for a visit.",
  "format": "social_caption",
  "language": "hinglish",
  "held_for_review": false,
  "summary": "",
  "findings": [],
  "facts_used": [
    {
      "id": "f-102",
      "category": "pricing",
      "label": "2BHK deposit"
    }
  ]
}

The same request, but the model added a price the brief and facts don't support.

{
  "brief": "New 2BHK listing",
  "format": "social_caption",
  "use_knowledge_base": false
}

{
  "copy": "2BHK ready to move in — only ₹35,000 deposit, limited time!",
  "held_for_review": true,
  "summary": "1 unsupported specific found.",
  "findings": [
    {
      "kind": "unsupported_specific",
      "category": "price",
      "text": "₹35,000",
      "reason": "This price does not appear in the brief, the source material, or your recorded facts."
    }
  ]
}

Shorten existing copy — the guard reports what the shortened version dropped.

{
  "text": "2BHK ready to move in, deposit ₹50,000, near Kandivali station.",
  "action": "shorten"
}

{
  "copy": "2BHK ready to move in near Kandivali station.",
  "action": "shorten",
  "changes": {
    "dropped_specifics": [
      "₹50,000"
    ],
    "added": [],
    "removed": [
      "deposit",
      "₹50,000"
    ]
  },
  "held_for_review": false
}

The seven formats with their real limits, the eleven languages, and the six rewrite actions.

{
  "formats": [
    {
      "value": "sms",
      "description": "an SMS under 160 characters",
      "max_chars": 160
    },
    {
      "value": "social_caption",
      "description": "a social caption with a hook in the first line, under 50 words",
      "max_words": 50
    }
  ],
  "languages": [
    "hinglish",
    "english",
    "hindi",
    "marathi",
    "..."
  ],
  "max_variations": 5
}

Run the guard over copy you wrote yourself. Needs no AI model.

{
  "text": "Guaranteed returns, no risk — invest today!",
  "use_knowledge_base": false
}

{
  "held_for_review": true,
  "summary": "2 regulated claims found.",
  "findings": [
    {
      "kind": "regulated_claim",
      "category": "financial",
      "text": "Guaranteed returns",
      "reason": "Reads as a financial guarantee."
    },
    {
      "kind": "regulated_claim",
      "category": "financial",
      "text": "no risk",
      "reason": "Reads as a financial guarantee."
    }
  ]
}

Keep a piece of copy, with its history and its flags, starting from a written draft.

{
  "content": "2BHK ready to move in — ₹35,000 deposit only.",
  "title": "Kandivali listing",
  "format": "social_caption"
}

{
  "id": "doc-9",
  "title": "Kandivali listing",
  "status": "in_review",
  "current_version_number": 1,
  "version": {
    "version_number": 1,
    "held_for_review": true
  },
  "flags": [
    {
      "id": "flag-1",
      "kind": "unsupported_specific",
      "category": "price",
      "text": "₹35,000",
      "status": "open"
    }
  ]
}

Try to approve — refused while a flag on the current version is still open.

{
  "status": "approved"
}

{
  "detail": {
    "message": "1 flag(s) on this version still need a decision. A person has to look at each one before this can be approved.",
    "open_flags": 1,
    "document_id": "doc-9"
  }
}

Decide the flag. Only after every open flag is decided can the document be approved.

{
  "decision": "rejected",
  "note": "Real deposit is ₹50,000 — draft needs a fix, not an override."
}

{
  "id": "flag-1",
  "status": "rejected",
  "reviewer_user_id": "usr_9",
  "open_flags_remaining": 0
}

What changed between two versions, and what specifically went missing.

{
  "changes": {
    "added": [
      "fixed"
    ],
    "removed": [],
    "dropped_specifics": []
  }
}

Localise, holding numbers and proper nouns fixed.

{
  "text": "2BHK ready to move in, deposit ₹50,000.",
  "language": "tamil"
}

{
  "copy": "2BHK உடனடியாக குடியேற தயார், டெபாசிட் ₹50,000.",
  "language": "tamil",
  "changes": {
    "dropped_specifics": []
  },
  "held_for_review": false
}