Pdf Chatbot

Study tools → ❓ Quiz → Generate quiz

Quick use cases

• Exam revision — upload lecture notes + slides, ask "explain X", get answers with page-numbered citations you can verify against the source.

• Clause lookup — find 7(c) in a contract or policy without scrolling; the tokenizer keeps sub-clause identifiers intact so the right section surfaces first.

• Onboarding Q&A — drop your handbook and runbooks in; new hires ask instead of interrupting, and every answer links back to the doc.

• Research triage — load 20 papers, ask which ones cover your method, get per-paper grounded summaries with traceable sources.

• Compliance evidence — answers are marked grounded / partially grounded / ungrounded, so you know when to trust one and when to open the PDF.

• Study-tool generation — flashcards, quiz questions and summaries built from your own material, not from the model's memory.

• Private/air-gapped deployment — self-hosted Postgres + pgvector, no document ever leaves your machine.

Updates:
Search speed — 35x faster

• Keyword search checked every passage in the library for every question. It now jumps straight to passages containing the search terms.

• Over 32,000 passages: 44ms → 1.25ms per query. The gain grows as the library grows — this scales, rather than being tuned.

Upload processing — 160x faster

• Duplicate-text removal compared every paragraph against every other. On a long book, 12.5 seconds. Now grouped so only plausible matches are compared: 0.08 seconds.

Both exact, not approximations

• Speed here usually costs accuracy. Neither does. The duplicate check is provably complete by a mathematical guarantee, enforced by an assertion in the code.

• Verified by running old and new side by side: scores identical, results identical, at every size tested.

Upload speed

• Large uploads ~4x faster. The step converting text for search ran one batch at a time, waiting on each before starting the next. A 10,000-passage document dropped from ~20 seconds of waiting to ~5.

• Hosted processing only. On-device setups were deliberately left alone — parallel work would slow them, not speed them.

Query efficiency

• Removed a redundant database call on every search. Ten searches now cost one lookup, not ten. Lower load and latency, better headroom as the corpus grows.

Risk controlled

• The parallel change had one dangerous failure mode: passages silently matched to the wrong search data — the system would look fine while every answer was wrong. We built a test that forces that condition and verifies all 500 pairings.

Assurance

• 260 automated checks green, up from 255

• Answer quality re-measured: unchanged

• Validated end-to-end

Updates:
Accuracy

• Found a defect in how the product indexes legal and technical references. "Article 7(c)" was stored as just "Article" — the clause letter discarded — so 7(a), 7(b) and 7(c) were indistinguishable. Any query naming a clause was effectively guessing.

• Now fixed. Every accuracy measure improved; the toughest — first result correct — rose from 91% to 96%.

• Matters most for regulatory, contract and compliance work, where the clause number is the question.

Infrastructure

• Switched the search index to a better-suited type: removes a fragile step that rebuilt it after every upload, and improves result quality.

Faster

• Uploading documents is ~3x quicker — a 200-page file now processes in a third of the time

• Study tools (summary, flashcards, quiz) start sooner: 5 trips to the AI service became 1

• Answers finish faster — fact-checking and follow-up suggestions now run at the same time instead of one after the other

Cheaper

• Our cost-saving feature was silently doing nothing (applied to the wrong part of each request). Now live — document excerpts are paid for once per session, not 3–4x per question

Confidence

• 245 automated tests passing

• Answer quality re-measured: identical — no quality traded for speed

• Verified end-to-end in a real browser
  

Fixed bugs

• MMR discarded fusion scores → recall@5 0.864 → 1.000, misses 3 → 0

• SESSION_TTL_MINUTES parsed but never enforced → sessions now expire

• "Clear conversation" left transcripts in Postgres → now clears

• Embeddings had worse error messages than chat for the same failure

Added

• Hybrid BM25 + vector retrieval, RRF fusion, MMR, optional reranker

• Citations to file+page, fake [S7] markers stripped, self-check verdicts

• Eval harness (recall@k/MRR/nDCG) — it caught the MMR bug

• Postgres/pgvector, auth, rate limiting, Docker, study tools

PDF Chatbot — Now Supports BYOK

Bring Your Own Key: connect your own OpenAI or Anthropic API key and use the PDF study assistant without relying on a shared platform key.

PDF Study Assistant

• Answers only from your PDFs — refuses when they don't cover it

• Every claim cited to file + page; self-check pass verifies claims against sources (✅/⚠️/🚩); fake citations stripped

• Hybrid BM25 + vector search, RRF fusion, MMR diversification, optional cross-encoder reranker

• Smart PDF prep: de-hyphenation, header stripping, structure-aware chunking, dedup

• Study tools: summaries, glossaries, Anki flashcards, scored quizzes; Markdown export

• Streamlit UI; Docker + Postgres/pgvector; Anthropic or OpenAI; fully offline mode

• Caching so re-uploads are free; password auth, rate limits, live cost tracking

• Eval harness (recall@k/MRR) with guards against fooling yourself; 205 tests, no network calls