Ask Lutecium Documentary Synthesis — Technical Design

Ask Lutecium Documentary Synthesis — Technical Design

This page describes the technical architecture of the documentary-synthesis component of Ask Lutecium. It is intended for software designers, researchers and technically interested readers.

Ask Lutecium deliberately separates semantic retrieval from language-model processing. The language model is not allowed to answer freely from its general knowledge. Its role is constrained to analysing material retrieved from Lutecium and selecting documentary propositions that can be verified against those sources.

1. General architecture

User question
     |
     v
Semantic retrieval
     |
     v
Relevant Mathèmes passages
     |
     v
Answerability analysis
     |
     v
Relevance classification
     |
     v
Documentary proposition extraction
     |
     v
Mechanical exact-quotation verification
     |
     v
Selection and ordering
     |
     v
Deterministic citation rendering
     |
     v
Displayed documentary synthesis

The semantic search and the documentary synthesis are distinct operations. Semantic search may search the Lutecium website, the Mathèmes de Lacan and mirror archives. At present, documentary synthesis is restricted to the Mathèmes de Lacan corpus.

2. Semantic retrieval

The question the user formulated is first transformed into a semantic representation using a multilingual embedding model. Similar passages are retrieved from the indexed Lutecium collections using vector similarity search constructed in a 384 dimensions vector.

This stage does not attempt to answer the question. It only identifies passages that are semantically close to the request.

This distinction is important: semantic similarity is not considered evidence that a passage answers the question.

3. Language-model processing

When the primary GPU synthesis server is available, the retrieved Mathèmes passages are submitted to several constrained language-model stages.

Answerability

The first stage determines whether the retrieved documentary material actually permits an answer. A thematic relationship alone is insufficient.

Relevance classification

Every retrieved passage is classified individually. Passages that are merely associated with the general topic, but do not directly contribute to answering the question, are excluded from synthesis.

Documentary proposition extraction

The model identifies propositions that are explicitly supported by the retained passages and associates each proposition with an exact quotation and a source identifier.

Mechanical verification

The decisive verification is performed by ordinary program code, not by the language model.

For every proposed support quotation, the program verifies that the exact character sequence is actually present in the source passage identified by the model. A quotation that cannot be found literally is rejected.

Ordering

After verification, the language model may select, group and order the validated documentary propositions. It is not permitted to invent additional theoretical transitions or explanatory sentences between them.

Deterministic rendering

Source numbers and the final presentation are generated programmatically. The model does not manufacture citations.

4. Seven anti-hallucination rules

No rules or weak rules leads these systems to produce and forge hallucinations. To prevent this to happen a series of seven rules are applied, and by being mandatory they minimize the hallucinations risk.

  1. Source-only rule. The language model may use only the passages supplied by the Lutecium retrieval system. General knowledge contained in the model is not accepted as documentary evidence.
  2. Answerability rule. Semantic proximity is not sufficient. The retrieved material must contain explicit propositions capable of answering the question. If it does not, the system must report that the documentary material is insufficient.
  3. Direct-relevance rule. Every source retained for synthesis must bear directly on the question. Passages that are merely thematically related must be discarded.
  4. Exact-quotation rule. Every documentary proposition must be associated with an exact quotation from an identified source.
  5. Mechanical-verification rule. The existence of every support quotation is checked mechanically against the retrieved source text. Verification is therefore independent of the language model assertion that the quotation exists.
  6. Immutable-document rule. Once an exact support has been verified, the documentary content displayed to the reader is constrained by that verified source material. Model-generated paraphrase is not allowed to substitute unsupported wording for the documentary proposition.
  7. No-generative-linking rule. The model may select, group and order validated documentary propositions, but it may not create causal relations, theoretical deductions, explanatory transitions or citations that are absent from the source material.

These rules substantially constrain the usual generative behaviour of a language model. They are designed specifically to prevent unsupported statements from entering the documentary synthesis. They should not, however, be interpreted as a theoretical guarantee that any language-model system is incapable of error.

5. Structured communication with the LLM

The synthesis stages communicate with the language model using constrained JSON structures rather than free-form prose.

The service validates the returned JSON before using it. If an LLM response is truncated or contains malformed JSON, the stage may be retried once with a larger output allowance. If valid structured output still cannot be obtained, the full synthesis pipeline is considered to have failed.

6. Primary GPU synthesis

The normal synthesis path uses a dedicated GPU inference server connected to the private Lutecium network.

The current primary server is run on a machine with a fast GPU which is kindly lent, running a 27-billion-parameter Qwen model through llama.cpp. The synthesis service itself runs on a Lutecium machine and dispatches inference requests to the primary GPU while retaining retrieval, verification and rendering under Lutecium control.

7. Degraded CPU mode

If the primary GPU server is unavailable, Ask Lutecium does not suspend semantic search.

The semantic results are returned immediately, while an asynchronous degraded synthesis may be started on the Lutecium machine using a smaller CPU-based language model.

The degraded mode consolidates the analysis into a single constrained language-model pass, followed by mechanical quotation verification and deterministic rendering.

This mode is slower, normally requiring approximately one to four minutes depending on the question, the retrieved passages and prompt-cache reuse.

Identical outstanding degraded requests are deduplicated so that repeated page reloads do not generate multiple expensive CPU jobs. The degraded queue is also bounded.

8. Failure handling

  • Primary GPU available and full synthesis succeeds: the normal documentary synthesis is displayed.
  • Primary GPU unavailable: semantic search is displayed immediately and asynchronous degraded synthesis is started.
  • Primary GPU available but a particular full synthesis fails: the request is transferred to asynchronous degraded synthesis rather than presenting a dead-end error to the user.

9. Separation between retrieval and generation

The semantic-search system determines which documents may be relevant. The language model then performs constrained documentary analysis. Finally, ordinary deterministic software verifies quotations and constructs citations.

The language model therefore does not act as an autonomous source of knowledge. It acts as a controlled analytical component between retrieval and deterministic verification.

10. Current scope and limitations

Documentary synthesis currently operates on the Mathèmes de Lacan corpus. Semantic search itself covers a broader collection including the Lutecium website and mirror archives.

The displayed synthesis should consequently be understood as a structured documentary view of the retrieved Mathèmes, not as an exhaustive theoretical interpretation of Lacan and not as a substitute for consulting the original texts.