HealthLog/src/lib/documents/content-index.ts

/**
 * v1.27.22 (Document vault P2) — the blind content-search index.
 *
 * Content search matches INSIDE a stored document without ever keeping its body
 * readable at rest. Two derived artefacts live in `DocumentContentIndex`:
 *
 *   1. `textEncrypted` — the normalised extracted text, AES-256-GCM at rest via
 *      the shared `encrypt()`-string-as-UTF-8 Bytes codec (same posture as the
 *      Coach columns). Recoverable server-side so key rotation can re-tokenise.
 *   2. `searchTokens` — each unique normalised token HMAC-SHA256'd under an
 *      HKDF-derived index subkey, truncated to a short hex tag. Deterministic
 *      (the same token always hashes to the same tag) so a query can be hashed
 *      the same way and matched with a GIN-accelerated array-overlap; one-way
 *      (opaque without the server-held subkey), so nothing readable is stored.
 *
 * The index subkey is HKDF-derived from the ACTIVE encryption key (P2-D7), never
 * the raw master key or the HMAC auth key. Because it follows the active key,
 * rotating the master key changes the subkey — the rotation script re-tokenises
 * every row from the decrypted `textEncrypted` so search keeps matching.
 *
 * Honest limit (P2-D6): the token index does WHOLE-TOKEN equality only — no
 * substring, prefix, or fuzzy matching. The vault list keeps the title/filename
 * ILIKE for substring on the short fields and unions the two result sets.
 */
import { createHmac } from "node:crypto";

import { decryptFromBytes, encryptToBytes } from "@/lib/ai/coach/bytes-codec";
import { deriveSubkey } from "@/lib/crypto";
import { prisma } from "@/lib/db";
import { expandQueryTokens } from "@/lib/documents/search-synonyms";

/**
 * Tokeniser algorithm version. Bump when the normalisation / tokenisation rules
 * change so a re-index can be detected unambiguously; it is orthogonal to the
 * index subkey (which follows the encryption key, not this version).
 */
export const CONTENT_TOKENIZER_VERSION = "1";

/** HKDF domain-separation label for the index HMAC subkey (P2-D7). */
const INDEX_SUBKEY_INFO = "healthlog:document-content-index:v1";

/** Minimum token length kept — drop 1-2 char noise ("mg", "x"). */
const MIN_TOKEN_LENGTH = 3;

/** Maximum token length kept — clamp a runaway OCR mash-up. */
const MAX_TOKEN_LENGTH = 64;

/** Cap on unique tokens stored per document — bounds the array + GIN entry. */
export const MAX_TOKENS_PER_DOCUMENT = 2048;

/**
 * Cap on the normalised text kept at rest per document (~64 KiB). The full body
 * is only needed to re-tokenise on rotation; a bounded prefix keeps the sibling
 * row lean and the token set representative without hoarding an unbounded blob.
 */
export const MAX_INDEX_TEXT_BYTES = 64 * 1024;

/**
 * v1.27.33 (Document vault P4) — cap on the VERBATIM text kept at rest per
 * document. Held at the same ~64 KiB budget as the normalised text so the whole
 * document still fits a mainstream chat context window when fed to the document
 * chat (≈16–20k tokens + the system prompt + history). The verbatim text is the
 * raw transcription (casing, accents, section names intact) so a chat can cite
 * the document faithfully; it is byte-bounded but NEVER lowercased or de-accented.
 */
export const MAX_VERBATIM_TEXT_BYTES = 64 * 1024;

/** Cap on query tokens hashed for a search (a search box, not a corpus). */
const MAX_QUERY_TOKENS = 24;

/**
 * A compact de/en stopword set. Dropping the highest-frequency function words
 * keeps the token array (and the GIN index) focused on content terms; a missed
 * stopword only wastes a slot, never a correctness bug.
 */
const STOPWORDS: ReadonlySet<string> = new Set([
  // English
  "the",
  "and",
  "for",
  "are",
  "was",
  "with",
  "that",
  "this",
  "have",
  "has",
  "not",
  "from",
  "you",
  "your",
  "his",
  "her",
  "she",
  "him",
  "they",
  "them",
  "were",
  "which",
  "who",
  "whom",
  "what",
  "when",
  "where",
  "will",
  "would",
  "there",
  "their",
  "been",
  "than",
  "then",
  "into",
  "out",
  "off",
  "per",
  // German
  "und",
  "der",
  "die",
  "das",
  "den",
  "dem",
  "des",
  "ein",
  "eine",
  "einer",
  "eines",
  "einem",
  "einen",
  "ist",
  "sind",
  "war",
  "waren",
  "wird",
  "werden",
  "nicht",
  "auch",
  "auf",
  "aus",
  "bei",
  "mit",
  "nach",
  "von",
  "vom",
  "zur",
  "zum",
  "für",
  "durch",
  "oder",
  "aber",
  "dass",
  "wie",
  "wenn",
  "sich",
  "als",
  "sowie",
  "wurde",
  "wurden",
  "bzw",
  "ggf",
]);

/**
 * Normalise raw extracted text: lowercase, strip diacritics (NFKD →
 * combining-mark removal), and cap to the storage budget. The cap is applied on
 * the byte length so the `textEncrypted` payload stays bounded regardless of
 * multi-byte content.
 */
export function normaliseIndexText(raw: string): string {
  const lowered = raw
    .normalize("NFKD")
    .replace(/\p{Diacritic}/gu, "")
    .toLowerCase();
  // Byte-bounded truncation — slice by chars until the UTF-8 length fits.
  if (Buffer.byteLength(lowered, "utf8") <= MAX_INDEX_TEXT_BYTES)
    return lowered;
  let out = lowered.slice(0, MAX_INDEX_TEXT_BYTES);
  while (Buffer.byteLength(out, "utf8") > MAX_INDEX_TEXT_BYTES) {
    out = out.slice(0, -256);
  }
  return out;
}

/**
 * Split normalised text into the deduped set of index tokens: alphanumeric
 * runs, length-gated, stopword-dropped, unique, capped. Insertion order is
 * preserved so the cap keeps the earliest-seen terms.
 */
export function tokenise(text: string): string[] {
  const normalised = normaliseIndexText(text);
  const seen = new Set<string>();
  // Split on anything that is not a letter or digit (Unicode-aware).
  for (const rawToken of normalised.split(/[^\p{L}\p{N}]+/u)) {
    if (seen.size >= MAX_TOKENS_PER_DOCUMENT) break;
    const token = rawToken.trim();
    if (token.length < MIN_TOKEN_LENGTH || token.length > MAX_TOKEN_LENGTH) {
      continue;
    }
    if (STOPWORDS.has(token)) continue;
    seen.add(token);
  }
  return [...seen];
}

/**
 * v1.27.33 (Document vault P4) — byte-bound the VERBATIM text for storage,
 * WITHOUT the normalisation `normaliseIndexText` applies. Casing, accents,
 * section names, and units survive intact so a document chat can cite the
 * document faithfully; only the byte length is clamped to the storage budget.
 */
export function captureVerbatimText(raw: string): string {
  if (Buffer.byteLength(raw, "utf8") <= MAX_VERBATIM_TEXT_BYTES) return raw;
  let out = raw.slice(0, MAX_VERBATIM_TEXT_BYTES);
  while (Buffer.byteLength(out, "utf8") > MAX_VERBATIM_TEXT_BYTES) {
    out = out.slice(0, -256);
  }
  return out;
}

/** The HKDF-derived HMAC subkey for the blind index. Never persisted or logged. */
function indexSubkey(): Buffer {
  return deriveSubkey(INDEX_SUBKEY_INFO);
}

/** HMAC-SHA256 one token under the index subkey; truncated to a 16-char hex tag. */
function hashToken(token: string, subkey: Buffer): string {
  return createHmac("sha256", subkey)
    .update(token, "utf8")
    .digest("hex")
    .slice(0, 16);
}

/**
 * Turn extracted text into the deduped array of blind token hashes to persist.
 * Deterministic under the active index subkey; opaque without it.
 */
export function tokeniseAndHash(text: string): string[] {
  const subkey = indexSubkey();
  const tokens = tokenise(text);
  const hashes = new Set<string>();
  for (const token of tokens) hashes.add(hashToken(token, subkey));
  return [...hashes];
}

/**
 * Hash a user's search query into the token tags to match against the index.
 * Same normalisation + subkey as `tokeniseAndHash`, so a whole-word hit in the
 * body produces the same tag. Returns [] when the query has no indexable token
 * (the caller then skips the content-match branch).
 */
export function hashQueryTokens(query: string): string[] {
  const subkey = indexSubkey();
  // v1.37.20 — expand through the curated German/English medical synonym
  // groups (`search-synonyms.ts`) BEFORE hashing, so "Blutdruck" also
  // matches a letter that says "blood pressure". Query-time only: the
  // stored index and the tokenizer version are untouched, and the expansion
  // happens before the one-way hash, so nothing readable moves anywhere new.
  const tokens = expandQueryTokens(tokenise(query).slice(0, MAX_QUERY_TOKENS));
  const hashes = new Set<string>();
  for (const token of tokens) hashes.add(hashToken(token, subkey));
  return [...hashes];
}

/** Encrypt normalised index text into the `Bytes` payload the schema stores. */
export function encryptIndexText(text: string): Uint8Array<ArrayBuffer> {
  return encryptToBytes(text);
}

/** Decrypt a stored index-text payload back to its plaintext. Throws on bad key. */
export function decryptIndexText(buf: Uint8Array): string {
  return decryptFromBytes(buf);
}

/**
 * v1.27.33 (Document vault P4) — encrypt the verbatim document text into the
 * `Bytes` payload the schema stores. Same AES-256-GCM codec as the normalised
 * index text; the only difference is the plaintext is un-normalised.
 */
export function encryptVerbatimText(text: string): Uint8Array<ArrayBuffer> {
  return encryptToBytes(text);
}

/** Decrypt a stored verbatim-text payload back to its plaintext. Throws on bad key. */
export function decryptVerbatimText(buf: Uint8Array): string {
  return decryptFromBytes(buf);
}

/** The document text a chat grounds on, plus which artefact it came from. */
export interface DocumentChatContext {
  /** The decrypted document text — verbatim when available, else normalised. */
  text: string;
  /** `"verbatim"` = the raw-fidelity capture; `"normalised"` = pre-P4 fallback. */
  source: "verbatim" | "normalised";
}

/**
 * v1.27.33 (Document vault P4) — load the best available document text for the
 * chat, owner-scoped. Prefers the verbatim capture (raw casing/accents/section
 * names → faithful citation); falls back to the normalised search text for a
 * row indexed before P4 that carries no verbatim column yet. Returns null when
 * the document has no content index at all — the chat is available ONLY for an
 * indexed document (the route maps this to a 422 "index first"). Decrypt failures
 * throw (fail-closed), exactly like every other `*Encrypted` read.
 */
export async function loadDocumentChatText(
  userId: string,
  documentId: string,
): Promise<DocumentChatContext | null> {
  const row = await prisma.documentContentIndex.findFirst({
    where: { documentId, userId },
    select: { textEncrypted: true, verbatimTextEncrypted: true },
  });
  if (!row) return null;
  if (row.verbatimTextEncrypted && row.verbatimTextEncrypted.byteLength > 0) {
    return {
      text: decryptVerbatimText(row.verbatimTextEncrypted),
      source: "verbatim",
    };
  }
  return { text: decryptIndexText(row.textEncrypted), source: "normalised" };
}

/**
 * Provenance of the indexed text.
 *   - `vision`    → an AI provider transcribed the stored original (image or PDF,
 *                   incl. scanned) — the AI-first primary path.
 *   - `text-ocr`  → browser-OCR text posted by the client (opt-in local OCR).
 *   - `local-pdf` → server-side text-layer extraction of a PDF (`pdf-parse`), no
 *                   provider, no egress — the fallback when no provider is usable.
 *   - `local-ocr` → server-side OCR of a scanned image/PDF (deferred follow-up).
 */
export type ContentIndexSource =
  "vision" | "text-ocr" | "local-pdf" | "local-ocr";

export interface UpsertContentIndexArgs {
  userId: string;
  documentId: string;
  /** Raw extracted text (normalised + capped inside). */
  text: string;
  source: ContentIndexSource;
  /** Provider type that produced the text (null on the text-ocr path). */
  providerType?: string | null;
}

/**
 * Populate or refresh one document's content index. Normalises + caps the text,
 * encrypts it, tokenises + hashes it, and upserts the 1:1 sibling row. Storing
 * only ciphertext + opaque hashes preserves the vault's encryption-at-rest
 * promise (A4). Idempotent — a re-index overwrites both artefacts in place.
 */
export async function upsertContentIndex(
  args: UpsertContentIndexArgs,
): Promise<{ tokenCount: number }> {
  const normalised = normaliseIndexText(args.text);
  const textEncrypted = encryptIndexText(normalised);
  const searchTokens = tokeniseAndHash(normalised);
  // v1.27.33 (Document vault P4) — additionally capture the VERBATIM text (raw
  // casing/accents, byte-capped only) for faithful citation in the document
  // chat. Derived from the same raw `text`, before normalisation.
  const verbatimTextEncrypted = encryptVerbatimText(
    captureVerbatimText(args.text),
  );
  await prisma.documentContentIndex.upsert({
    where: { documentId: args.documentId },
    create: {
      documentId: args.documentId,
      userId: args.userId,
      textEncrypted,
      verbatimTextEncrypted,
      searchTokens,
      source: args.source,
      providerType: args.providerType ?? null,
      tokenizerVersion: CONTENT_TOKENIZER_VERSION,
    },
    update: {
      textEncrypted,
      verbatimTextEncrypted,
      searchTokens,
      source: args.source,
      providerType: args.providerType ?? null,
      tokenizerVersion: CONTENT_TOKENIZER_VERSION,
    },
  });
  return { tokenCount: searchTokens.length };
}