blog_rewriter Lambda Overview

Triggered by messages on the SQS blog-rewriter queue, this Lambda rewrites the body of a past blog (old_blog) without changing its content. When the blog is tied to a race, the race result, race video and course information (horse racing only) are taken into account as well.

The rewritten body is passed to the blog_builder Lambda over the SQS blog-builder queue, which renders it to HTML and stores it in the blog collection.

Trigger

Sender	When
App Lambda (`odds_poc_app`)	On a rewrite request (`POST /v1/old_blogs/{old_blog_id}/rewrite`)
App Lambda (`odds_poc_app`)	On a blog generation request (`POST /v1/blogs/generation`, via the `blog-rewrite` queue)

SQS event payload (input)

{
  "blog_id": "string",
  "old_blog_id": "string",
  "race_id": "string",
  "racing_type": "horse_racing"
}

Field	Type	Required	Notes
blog_id	string	◯	`_id` of the `blog` row (in `generating` state)
old_blog_id	string	◯	`old_blog._id` being rewritten
race_id	string		`_id` of the linked race; optional, since not every past blog has one
racing_type	enum		`auto_racing` / `bicycle_racing` / `horse_racing`

Processing flow

Receive the SQS event and parse the payload
Connect to DocumentDB
Check the target blog is still status=generating; otherwise exit without calling Bedrock
Fetch the old_blog
If the blog is tied to a race
- Fetch the race result and build the full finishing-order table and the related-information race card
- Fetch the race video and course image (horse racing only) from S3
- Stage 1: analyse the race with race_analysis_agent
- If it is not tied to a race, this stage is skipped and the rewrite proceeds with "no race information"
Stage 2: rewrite the body in light of the analysis, structured as block JSON
Stage 3: generate the "10-second summary" from the rewritten body
Send to the SQS blog-builder queue

flowchart TD
    AppLambda([App Lambda]) -->|SQS| Start
    Start[Receive SQS trigger] --> DB[Connect to DocumentDB]
    DB --> Guard[Check blog is generating]
    Guard -->|not generating| Skip[Exit without work]
    Guard -->|OK| Old[Fetch old_blog]
    Old -->|race linked| Race[Fetch race result, build table and race card]
    Old -->|no race| NoRace[Continue with no race information]
    Race --> Media[Fetch video / course image from S3]
    Media --> Analyze[Stage 1: race_analysis_agent]
    Analyze --> Rewrite
    NoRace --> Rewrite[Stage 2: rewrite body into block JSON]
    Rewrite --> Summary[Stage 3: generate the 10-second summary]
    Summary --> SendSQS[Send to blog-builder queue]
    SendSQS -.->|SQS| Builder([blog_builder Lambda])

Rewriting policy

Do not change the content — preserve the original information and the author's voice
The shared tone-of-voice text from packages/tone_of_voice is injected into the prompt
The body is structured as an array of "blocks" matched to the content (see blog_builder for the block types)
Relative image src values are absolutized against the source article URL (old_blog.source_url)
- Body images of past blogs are served from the same host as the source article. Only when source_url is missing does it fall back to https://www.oddspark.com/
For race-linked blogs, entry-list enumerations are deleted rather than rewritten (entrant and finishing-order data is rendered separately from the official data)
- The prompt says so, and the code drops enumeration blocks as a safety net when the LLM does not comply
The colour of a prediction number badge is determined by the bracket number in horse racing. Using the horse number would give the wrong colour, so the bracket number is looked up from the race result by horse name; if it cannot be resolved, no colour is applied (no colour is more accurate than a wrong one)

Regeneration guards

Guard	What it checks	Limit
Block JSON	The output parses as block JSON. If it is broken or truncated by the token limit, the error is attached as a correction instruction and the text is rewritten	2 retries
NG words	The rewritten text contains no prohibited words	2 retries

Words from the source blog are excluded

A rewrite preserves the content, so words already in the source (e.g. a jockey's name) were not introduced by the AI and cannot be removed by regenerating. Words present in the original body are therefore excluded from screening.

Broken JSON is never published as HTML

If the model attempted block JSON and produced something broken, falling back to HTML would publish the raw JSON text as the article body. Instead, nothing is published: the error is attached as a correction instruction and the text is rewritten in-process — faster than waiting for SQS redelivery, and more likely to succeed because a correction can be supplied.

If the limit is reached without resolution, generation fails and nothing is published.

This Lambda does not generate the related-information section.

Concern	Owner
Attaching embeddings	odds_poc_app (on `POST /v1/old_blogs`, attached to `old_blog`)
Carrying vectors over	odds_poc_app (when creating the `blog` for a rewrite)
Generating related information	odds_poc_app (`POST /v1/blogs/{blog_id}/related_generate`)

Models

Use	Model
Race video / course image analysis (stage 1)	Amazon Nova Pro (supports video input)
Rewriting and summary (stages 2 and 3)	Claude Sonnet (Nova Pro when `BEDROCK_CLAUDE_SONNET_PROFILE_ARN` is unset)

The output token limit is 16384 on Claude and 8192 on Nova, because block JSON for a long column that gets truncated mid-output cannot be parsed.

Error handling

Situation	Behaviour
Failure on a non-final attempt	Re-raise and let SQS redelivery recover
Failure on the final attempt	Hard-delete the `status=generating` blog and finish without re-raising