Demo · 7 September 2026

TenderLens

HTML is split into chunks and indexed through a queue. Search runs in pgvector; answers link to the retrieved chunks. The demo has four test documents. Embeddings and the LLM are mocked; the external crawler is disabled.

Demo ↗Documentation →Source code ↗

Implementation history: since 20 August 2026. Public demo: 7 September 2026.Version a618d1e712 interface screenshots

01 / The task

Document search with FastAPI and pgvector

Search over a fixed document set. Answers need links to the source chunks. If there are no relevant matches, the response says so.

Use case: Searching and reading a document collection.

Work done

Document normalization, event-based indexing and chunk storage.
Search and answer API with source links. A separate response for empty context.
Test corpus, deterministic AI mocks and per-session limits.

Screenshots

Screens, forms, errors

Screenshots, diagrams and source evidence retain their original language; captions and descriptions are translated.

02 / Technology

Stack

FastAPI / Pydantic

HTTP contracts, validation and search responses.

NATS JetStream

Event delivery to the indexer and replay.

PostgreSQL / pgvector

Documents, attachments, chunks and cosine search.

SQLAlchemy / Python

Normalization, extraction, chunking and content_hash verification.

03 / Decisions

How the project works

Seed sends a real TenderChangedV1 event through NATS; the indexer extracts HTML and creates chunks.
PostgreSQL VECTOR(1024) performs cosine search. FakeAIProvider creates embeddings using the hashing trick.
Event replay uses content_hash: an unchanged document does not duplicate chunks.
Answers use an explicitly labeled template based on retrieved passages; the external crawler and Ollama are disabled.

04 / Verification

Demo workflows

Find a document about servers or an Angular portal.
Ask a question and open the answer's sources.
Enter a query outside the corpus and check the missing-context response.
Replay indexing events: the four documents and 16 chunks should remain without duplication.

Check results

12 meaningful screenshots of the actual interface retained.
Each screenshot includes its date, scenario and code revision.
Peak container RAM in the short test: 127.04 MiB. Shared services are counted separately; the shop includes its own database.
6 scenario repetitions measured; median 997.0 ms, maximum 2738 ms. Timing includes the browser and HTTPS; methodology and limitations are included.
Four ready documents, 16 chunks. Replaying four NATS events created no duplicates.

Limitations

Four synthetic documents and 16 chunks form a demonstration corpus, not a database of real procurement records.
Cosine scores and template answers are not an evaluation of a production model's quality.
Up to 30 requests per minute per session and 1,000 active demo keys. User queries are not stored.
Worker health confirms the connection at startup; a separate stalled-computation detector is not implemented.

Technical notes

Modules, API, data model, setup and check results.

Documentation →Materials ↓