AI Pipeline¶

diff·log uses a multi-step AI pipeline to generate personalized diffs. Each step can be powered by a different provider, and all API calls are made directly from the browser using the user's own keys.

Providers¶

Users configure API keys for one or more providers, then assign a provider to each pipeline step. Curation and synthesis require a provider; search is optional.

Overview¶

graph LR
    subgraph Inputs
        P[Profile]
        F[Feeds API]
    end

    subgraph "AI Pipeline"
        H1[Curation: Resolve Sources]
        H2[Curation: Filter Feeds]
        S1[Search: Web Search]
        S2[Synthesis: Generate Diff]
    end

    subgraph "On-Demand"
        T[TLDR: Summarize Article]
    end

    P <--> H1
    H1 --> F
    F --> H2
    P --> S1
    H2 --> S2
    S1 --> S2
    S2 --> D[Diff]
    D -.->|user clicks| T

Pipeline Steps¶

1. Source Resolution (curation provider)¶

When users add custom languages, frameworks, tools, or topics not in the predefined mappings, the curation provider resolves them to feed sources.

Predefined mappings exist in src/lib/utils/feeds.ts:

• LANGUAGE_SUBREDDITS — e.g., TypeScript → ['typescript']
• FRAMEWORK_SUBREDDITS — e.g., React → ['reactjs']
• TOOL_SUBREDDITS — e.g., Docker → ['docker']
• TOPIC_SUBREDDITS — e.g., AI/ML & LLMs → ['MachineLearning', 'LocalLLaMA']

Custom items (anything not in these maps) are resolved via resolveSourcesForItem():

// Returns structured sources via JSON
{
  subreddits: ['homelab', 'selfhosted'],
  lobstersTags: ['unix', 'devops'],
  devtoTags: ['homelab', 'selfhosted']
}

Resolved mappings are cached in profile.resolvedMappings to avoid repeated API calls.

Provider fallback for curation tasks: DeepSeek → Gemini → Anthropic Haiku (cheapest available is preferred).

2. Feed Curation (curation provider)¶

General feeds (Hacker News, Lobsters) contain broad tech news. Before including them in the prompt, the curation provider filters for relevance to the user's profile.

Process (curateGeneralFeeds() in src/lib/utils/feeds.ts):

1. Format all HN + Lobsters items with indices
2. Send to curation provider with profile context
3. Provider returns indices of relevant items
4. Only selected items are included in the prompt

This reduces noise while keeping important cross-cutting news (security alerts, major releases).

3. Web Search (search provider, optional)¶

The search provider finds recent news relevant to the user's profile.

Provider priority: Serper → Perplexity → Anthropic (cheapest first).

Process (searchWeb() in src/lib/utils/search.ts):

1. Build search context from profile (technologies + topics)
2. Execute searches via the selected provider
3. Extract results in structured format: title, URL, snippet
4. Format as WEB SEARCH RESULTS section in prompt

Web search runs in parallel with feed fetching for performance.

4. Diff Synthesis (synthesis provider)¶

The main generation combines all sources into a single prompt for the selected synthesis provider.

Anthropic uses structured output via tool_choice (submit_diff tool). Other providers use plain text output with system instructions and a title extracted from the first ## heading.

Prompt structure (buildPrompt() in src/lib/utils/prompt.ts):

FORMAT instructions (link formatting, score display)
PROFILE (name, technologies, topics, custom focus)
DEPTH instruction
WEB SEARCH RESULTS (if available)
REAL-TIME FEED DATA (curated items with URLs)
SECTION GUIDANCE (suggested sections with emoji prefixes)
PREVIOUS DIFF (to avoid repetition)

Output:

• title — Short creative title (3-8 words)
• content — Full markdown starting with date line
• cost — Estimated cost based on token usage

5. TLDR Summaries (curation provider, on-demand)¶

When a user clicks the TLDR button on a paragraph, the app fetches the linked article and generates a context-aware summary using the curation provider chain.

Article fetching (fetchArticleText() in src/lib/utils/tldr.ts):

1. Try Jina Reader — free extraction service that returns clean text/markdown
2. Fall back to /api/fetch-article — server-side fetch via Cloudflare worker that strips HTML
3. Validate content is real (min 200 chars, no CAPTCHA/login markers)
4. Truncate to ~4000 chars to keep LLM costs low

Summarization (summarizeArticle() in src/lib/utils/tldr.ts):

The prompt includes the paragraph's existing one-liner so the LLM adds depth rather than repeating it. The summary targets 4-6 short paragraphs with concrete details — version numbers, metrics, technical implications.

Caching: Summaries are stored in localStorage (difflog-tldrs) keyed by profileId → diffId:pIndex. Cached TLDRs show a persistent gold button. Not synced to the server — local-only, cheap to regenerate.

Cost: Same as curation tasks (uses completeJson with 768 max tokens). Roughly $0.0001-0.002 per TLDR depending on provider.

Depth Levels¶

Users select a reading depth that controls how detailed the generated diff will be:

The depth is passed to the generation prompt via DEPTH_INSTRUCTIONS in src/lib/utils/prompt.ts.

Intelligence Window¶

The time window adapts to when the user last generated a diff:

• No previous diff: 7 days
• Recent diff: Days since last diff (clamped 1-7)

This is displayed as "Intelligence Window: Past N days" in the diff header.

Cost Estimates¶

Costs vary significantly by provider selection. Estimated costs per diff at standard depth: