diff options
| author | Ralph Amissah <ralph.amissah@gmail.com> | 2026-09-24 16:33:53 -0400 |
|---|---|---|
| committer | Ralph Amissah <ralph.amissah@gmail.com> | 2026-09-24 16:33:53 -0400 |
| commit | bbe505c8d48e822c012c9820dfe93b83dc61bcab (patch) | |
| tree | 3385f5420959ff20af295c3125a080297555a72b /dlang/src/output_llm_jsonl.d | |
| parent | dealing with six properties including .is_of_type (diff) | |
--validate, and a second reader, in Gleam
--validate: the specification, executed. A validator written from
ssp-format-specification.md and from nothing else, so that a file can
be held against the document rather than against an implementation.
Each finding names the specification section it comes from, so a
disagreement can be taken back to the text.
It checks the version banner and its major part, the file shape and
the order of the preamble blocks and the nine sections, the two
spaces of a preamble pair, the object declaration and its sixteen
kinds and eight heading levels, that a property is one of the 42 and
in the emitted order, that text lines carry their prefix, the ten
keys of doc_has, and that no line is of a shape the specification
does not define.
Every .ssp of the markup document samples validates clean. Five
classes of deliberate breakage are caught: an unknown major version, a
property that is not one of the 42, a property out of order, a kind
that is not one of the sixteen, and a block that is neither preamble
nor section. Writing it found two rules the specification implied
and did not state (repeated properties are consecutive; no property
follows the text), and both were written into it.
One repository, a reader per language. The readers are not separate
tools that share a name. They are the same specification implemented
more than once, and the interesting thing about them is that they
agree, so they are published as one thing with that claim at the top
of it. The D reader moves to dlang/; the Gleam reader arrives at
gleam/. The language directory is named for the language and not
for the tool: gleam refuses a hyphenated package name, so ssp-reader
cannot be a gleam package and must be ssp_reader.
A .gitignore for the repository, denying everything and whitelisting
what belongs, covering both languages.
Both build under nix, sandboxed and without network. There is no gleam
builder in nixpkgs and gleam build fetches from Hex, so the Gleam build
uses nix-gleam's builder, kept in gleam/nix-gleam/ (MIT), which fetches
each Hex package by the checksum manifest.toml already carries, with no
hash to maintain; the result is an escript wrapped with erlang. The D
reader depends on nothing beyond phobos, so --skip-registry=all is the
whole of what it takes: ldc by default, dmd available as ssp-reader-dmd
(given gcc14, GCC 15 having introduced nullptr in headers dmd's ImportC
does not yet parse).
(assisted by Claude-Code)
Diffstat (limited to 'dlang/src/output_llm_jsonl.d')
| -rw-r--r-- | dlang/src/output_llm_jsonl.d | 121 |
1 files changed, 121 insertions, 0 deletions
diff --git a/dlang/src/output_llm_jsonl.d b/dlang/src/output_llm_jsonl.d new file mode 100644 index 0000000..3d34089 --- /dev/null +++ b/dlang/src/output_llm_jsonl.d @@ -0,0 +1,121 @@ +module output_llm_jsonl; + +import std.array : split, join, replace; +import std.conv : to; +import std.stdio : writeln, File; +import std.string : strip, indexOf, stripRight; +import ssp_types; +import output_llm_breadcrumbs : stripMarkup; + +/+ ↓ JSONL output with embeddings-ready chunks for RAG pipelines + one JSON object per line, each containing: + doc, ocn, path (breadcrumb), type, text (clean), + images (only where the object embeds one, so that a chunk whose + text reads around a figure carries the figure's identity), + index_terms (from bookindex entries associated with this object) ++/ +void outputLlmJsonl(SspDocument doc, string output_file) { + string[] lines; + string doc_id = doc.meta("title.main"); + string doc_author = doc.meta("creator.author"); + + /+ ↓ build bookindex -> OCN mapping for index_terms + bookindex objects in the .ssp have parent_ocn linking to body objects + but the index terms are in the text of bookindex objects, not + directly attached to body objects. We use a simpler approach: + headings carry heading_ancestors_text which gives context. +/ + + /+ ↓ track current heading path +/ + string current_path = ""; + + foreach (obj; doc.body) { + if (obj.is_a == "heading") { + current_path = buildPath(doc_id, obj.heading_ancestors_text, obj); + /+ ↓ emit heading as a chunk too +/ + string heading_text = stripMarkup(obj.text); + if (heading_text.strip.length > 0) { + lines ~= jsonLine(doc_id, doc_author, obj.ocn, current_path, "heading", heading_text, obj.images); + } + continue; + } + + /+ ↓ markup commentary is not content, and least of all is it a chunk + to embed: it would be indexed as though the author had written it + into the document + +/ + if (!obj.renders) { continue; } + + string text = stripMarkup(obj.text); + if (text.strip.length == 0) continue; + + lines ~= jsonLine(doc_id, doc_author, obj.ocn, current_path, obj.is_a, text, obj.images); + } + + /+ ↓ write output +/ + if (output_file.length > 0) { + auto f = File(output_file, "w"); + foreach (line; lines) f.writeln(line); + } else { + foreach (line; lines) writeln(line); + } +} + +private string jsonLine(string doc_id, string doc_author, int ocn, string path, string type, string text, SspImage[] images) { + return `{"doc":` ~ jsonStr(doc_id) + ~ `,"author":` ~ jsonStr(doc_author) + ~ `,"ocn":` ~ ocn.to!string + ~ `,"path":` ~ jsonStr(path) + ~ `,"type":` ~ jsonStr(type) + ~ `,"text":` ~ jsonStr(text) + ~ jsonImages(images) + ~ `}`; +} + +private string jsonImages(SspImage[] images) { + /+ ↓ omitted entirely where the object embeds no image +/ + if (images.length == 0) return ""; + string[] items; + foreach (img; images) { + string item = `{"name":` ~ jsonStr(img.name); + if (img.missing) { + item ~= `,"missing":true`; + } else { + item ~= `,"sha256":` ~ jsonStr(img.sha256) + ~ `,"bytes":` ~ img.bytes.to!string; + if (img.width > 0 && img.height > 0) { + item ~= `,"width":` ~ img.width.to!string + ~ `,"height":` ~ img.height.to!string; + } + } + items ~= item ~ `}`; + } + return `,"images":[` ~ items.join(",") ~ `]`; +} + +private string buildPath(string doc_title, string ancestors_text, SspObject obj) { + string[] parts; + if (ancestors_text.length > 0) { + foreach (part; ancestors_text.split("|")) { + string clean = stripMarkup(part).strip; + if (clean.length > 0 && clean != doc_title) { + parts ~= clean; + } + } + } + string heading_text = stripMarkup(obj.text).strip; + if (parts.length > 0) + return parts.join(" > ") ~ " > " ~ heading_text; + else + return heading_text; +} + +private string jsonStr(string s) { + /+ ↓ JSON string escaping +/ + string r = s; + r = r.replace(`\`, `\\`); + r = r.replace(`"`, `\"`); + r = r.replace("\n", `\n`); + r = r.replace("\r", `\r`); + r = r.replace("\t", `\t`); + return `"` ~ r ~ `"`; +} |
