Compare commits
3 commits
c1260cfa68
...
6acc587dd5
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
6acc587dd5 | ||
|
|
e45f1c04d0 | ||
|
|
aaf5618466 |
3 changed files with 73 additions and 13 deletions
|
|
@ -16,7 +16,7 @@ import { audit } from "../lib/audit";
|
||||||
import { callClaude } from "../lib/claude";
|
import { callClaude } from "../lib/claude";
|
||||||
import { env } from "../lib/env";
|
import { env } from "../lib/env";
|
||||||
import { query } from "../lib/pg";
|
import { query } from "../lib/pg";
|
||||||
import { hybridSearch, type SearchHit } from "../lib/search";
|
import { fetchDocChunks, hybridSearch, type SearchHit } from "../lib/search";
|
||||||
import { writeCaseReport } from "../tools/write_case_report";
|
import { writeCaseReport } from "../tools/write_case_report";
|
||||||
|
|
||||||
const HERE = path.dirname(fileURLToPath(import.meta.url));
|
const HERE = path.dirname(fileURLToPath(import.meta.url));
|
||||||
|
|
@ -286,18 +286,27 @@ export async function runCaseWriter(task: CaseWriterTask): Promise<
|
||||||
|
|
||||||
const filter = `%${topic.toLowerCase()}%`;
|
const filter = `%${topic.toLowerCase()}%`;
|
||||||
|
|
||||||
// Grounding pass — retrieve top scenes from the corpus via hybrid_search.
|
// Grounding pass — assemble the scenes the narrator weaves from.
|
||||||
// This is what gives the narrator real verbatim material to weave. Without
|
//
|
||||||
// this, the case-writer only sees pre-digested artefacts (which is what
|
// For a per-document case file (doc_id set) we pull THIS document's own
|
||||||
// produced the academic prose in v1).
|
// chunks in reading order. A hybridSearch keyed on the document's
|
||||||
const scenes = await hybridSearch({
|
// auto-derived topic ("Fbi Photo B20", "Doc 59 214434 …") returns zero
|
||||||
|
// hits even though the doc has dozens of embedded chunks — the dense gate
|
||||||
|
// rejects them all because the garbage topic has no semantic neighbours.
|
||||||
|
// That single bug skipped 61 of 75 batch documents on its own.
|
||||||
|
//
|
||||||
|
// For a corpus-wide topic report (no doc_id) the semantic search is
|
||||||
|
// exactly right — we want the strongest chunks across all documents.
|
||||||
|
const docIdFilter = task.doc_id ?? null;
|
||||||
|
const scenes = docIdFilter
|
||||||
|
? await fetchDocChunks(docIdFilter, lang, 24).catch(() => [] as SearchHit[])
|
||||||
|
: await hybridSearch({
|
||||||
query: topic, lang,
|
query: topic, lang,
|
||||||
doc_id: task.doc_id ?? null,
|
doc_id: null,
|
||||||
top_k: 18,
|
top_k: 18,
|
||||||
recall_k: 80,
|
recall_k: 80,
|
||||||
max_dense_dist: 0.55,
|
max_dense_dist: 0.55,
|
||||||
}).catch(() => [] as SearchHit[]);
|
}).catch(() => [] as SearchHit[]);
|
||||||
const docIdFilter = task.doc_id ?? null;
|
|
||||||
|
|
||||||
// Pull artefacts SEQUENTIALLY. The investigator role has rolconnlimit=4 and
|
// Pull artefacts SEQUENTIALLY. The investigator role has rolconnlimit=4 and
|
||||||
// pool.max=4; Promise.all of 5 queries × max_parallel=2 jobs would demand
|
// pool.max=4; Promise.all of 5 queries × max_parallel=2 jobs would demand
|
||||||
|
|
|
||||||
|
|
@ -67,6 +67,50 @@ export interface HybridSearchOpts {
|
||||||
max_dense_dist?: number;
|
max_dense_dist?: number;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* Fetch a single document's own chunks — no semantic gating. For a
|
||||||
|
* per-document case file the narrator wants THIS document's substance, not
|
||||||
|
* a corpus search; a hybridSearch keyed on the document's (often garbage)
|
||||||
|
* auto-derived topic returns zero hits even though the doc has dozens of
|
||||||
|
* embedded chunks.
|
||||||
|
*
|
||||||
|
* We pick the most substantive chunks (by content length, deprioritising
|
||||||
|
* pure redaction boxes) and THEN present them in reading order. Naively
|
||||||
|
* taking the first N by `order_global` starves the narrator on long files:
|
||||||
|
* the opening chunks of a 1000-chunk FBI dossier are cover pages, routing
|
||||||
|
* slips, classification stamps and redaction boxes — administrative front
|
||||||
|
* matter, not narrative. The substance sits deeper in the file, so a
|
||||||
|
* length-ranked pick surfaces it regardless of position, while the final
|
||||||
|
* reading-order sort keeps the story coherent.
|
||||||
|
*/
|
||||||
|
export async function fetchDocChunks(
|
||||||
|
doc_id: string,
|
||||||
|
_lang: "pt" | "en" = "pt",
|
||||||
|
limit = 24,
|
||||||
|
): Promise<SearchHit[]> {
|
||||||
|
if (!doc_id) return [];
|
||||||
|
return await query<SearchHit>(
|
||||||
|
`WITH ranked AS (
|
||||||
|
SELECT chunk_pk, doc_id, chunk_id, page, type, bbox,
|
||||||
|
content_en, content_pt, classification,
|
||||||
|
order_global, order_in_page,
|
||||||
|
length(COALESCE(content_en,'') || COALESCE(content_pt,'')) AS richness
|
||||||
|
FROM public.chunks
|
||||||
|
WHERE doc_id = $1
|
||||||
|
AND is_searchable = TRUE
|
||||||
|
AND length(COALESCE(content_en,'') || COALESCE(content_pt,'')) > 40
|
||||||
|
ORDER BY (type = 'redaction') ASC, richness DESC
|
||||||
|
LIMIT $2
|
||||||
|
)
|
||||||
|
SELECT chunk_pk, doc_id, chunk_id, page, type, bbox,
|
||||||
|
content_en, content_pt, classification,
|
||||||
|
1.0::float8 AS score, NULL::int AS bm25_rank, NULL::int AS dense_rank
|
||||||
|
FROM ranked
|
||||||
|
ORDER BY order_global ASC NULLS LAST, page ASC, order_in_page ASC`,
|
||||||
|
[doc_id, limit],
|
||||||
|
);
|
||||||
|
}
|
||||||
|
|
||||||
export async function hybridSearch(opts: HybridSearchOpts): Promise<SearchHit[]> {
|
export async function hybridSearch(opts: HybridSearchOpts): Promise<SearchHit[]> {
|
||||||
const {
|
const {
|
||||||
query: q,
|
query: q,
|
||||||
|
|
|
||||||
|
|
@ -81,7 +81,14 @@ export async function generateMetadata(
|
||||||
const c = await loadCase(slug);
|
const c = await loadCase(slug);
|
||||||
if (!c) return { title: "Case file not found" };
|
if (!c) return { title: "Case file not found" };
|
||||||
|
|
||||||
const title = locale === "pt-br" ? (c.fm.topic_pt_br ?? c.fm.topic ?? slug) : (c.fm.topic ?? slug);
|
// Prefer the narrator's body H1 (the magazine headline) over the generic
|
||||||
|
// auto-derived frontmatter topic ("Dow Uap D44 …"). This is what search
|
||||||
|
// engines and link cards show, so it must be the human title. Two H1s when
|
||||||
|
// present: EN then PT-BR; fall back across them, then to frontmatter.
|
||||||
|
const h1s = [...c.body.matchAll(/^#\s+(.+)$/gm)].map((m) => m[1].trim());
|
||||||
|
const title = locale === "pt-br"
|
||||||
|
? (h1s[1] ?? h1s[0] ?? c.fm.topic_pt_br ?? c.fm.topic ?? slug)
|
||||||
|
: (h1s[0] ?? c.fm.topic ?? slug);
|
||||||
const desc = pickLead(c.body, locale).slice(0, 200);
|
const desc = pickLead(c.body, locale).slice(0, 200);
|
||||||
const canonical = `${SITE_URL}/c/${slug}`;
|
const canonical = `${SITE_URL}/c/${slug}`;
|
||||||
// OG image — use the case's editorial illustration when present. WhatsApp,
|
// OG image — use the case's editorial illustration when present. WhatsApp,
|
||||||
|
|
|
||||||
Loading…
Reference in a new issue